Every week I get the same DM.

“Abhijay, every entry level job wants 2 years of experience. How do I get experience without experience?”

I know that feeling personally. Before Google said yes, I collected hundreds of rejections. The worst part wasn’t the no. It was the catch-22. You can’t get the job without experience, and you can’t get experience without the job.

This week, the biggest AI company in the world broke that loop. And they’re paying $85,000 for it.

What just happened

On June 11, Anthropic (the company behind Claude) announced Claude Corps. They’re putting $150 million into hiring 1,000 fellows and placing them inside nonprofits across the US to deploy AI.

Here’s what a fellow gets:

• Salary: $85,000 per year, plus full benefits

• Duration: 12 months, full time, in person at a host nonprofit

• Training: an intensive bootcamp on prompt design, building with the Claude API, and AI evaluations, then 5 hours of structured training every week

• Access: office hours with Anthropic’s own technical staff, mentorship from CodePath, and a generous Claude token budget

• Outcome: a year of real AI deployment experience on your resume, backed by the Anthropic name

Read that training list again. Prompt design. API development. AI evals. Companies are paying six figures for people with exactly those skills right now. Anthropic is paying YOU to learn them.

Here’s the part that made me sit up

Most programs like this filter for the most experienced applicants. This one does the opposite.

If you have more than 2 years of full time work experience, you can’t apply.

The full eligibility list:

• 18 or older

• Under 2 years of full time work experience

• No degree required. None.

• Authorized to work in the US

• Comfortable working with Claude

• Willing to relocate if needed (they provide support)

A quick note for my international readers, because I know half of you are asking already. The requirement is US work authorization. If you’re an F-1 student, OPT counts as work authorization, but a 12 month placement has specifics you need to confirm with your DSO before applying. Don’t skip that conversation. I learned the hard way that immigration paperwork doesn’t forgive assumptions.

The deadline is real

Applications for the first cohort (100 fellows, starting October 2026) close on July 17.

That’s five weeks from today. Later cohorts start in January 2027 and August 2027, but the first cohort is the smallest and the least competitive window you’ll ever get. Right now this is 24 hours old and most people haven’t heard of it.

Apply here: anthropic.com/claude-corps/fellow

The free prep stack (everything you need is $0)

Here’s everything you need, in the order I’d actually do it over the next five weeks. All of it is free. All of it comes straight from Anthropic. Bookmark this section.

Step 0: Read the source documents

Before anything else, read the two pages that decide whether you get in.

• The application (start here, confirm the deadline): anthropic.com/claude-corps/fellow

• The announcement page: anthropic.com/news/claude-corps

Read the announcement like a job description, because that’s what it is. Study the host organizations (YMCA, Goodwill, food banks, veteran groups). Notice the words they use about impact and capacity. Mirror that language in your application.

Step 1: Anthropic Academy (get the certificates)

Anthropic’s official learning platform. 17+ courses, all free, all with free certificates you can add to your LinkedIn and resume. Main catalog here: anthropic.skilljar.com.

Do these three, in order:

• Claude Platform 101: anthropic.skilljar.com/claude-platform-101

• Building with the Claude API (their biggest course, 84 lectures): anthropic.skilljar.com/claude-with-the-anthropic-api

• Claude Code in Action: anthropic.skilljar.com/claude-code-in-action

This is almost the exact curriculum the fellowship trains you on. Prompt design. API building. Evals. Walking in already knowing it is an unfair advantage.

Step 2: Anthropic’s AI Fluency courses on Coursera (the nonprofit angle)

Five free courses with certificates. Here’s the cheat code: the series includes a track aimed at nonprofit professionals, and Claude Corps places you inside nonprofits. Finishing that and naming it in your application is the most obvious signal you can send.

• Foundational course (about 3 hours): AI Fluency: Framework & Foundations

• Full Anthropic catalog on Coursera (find the nonprofit and student tracks here): coursera.org/partners/anthropic

Step 3: The hands-on tutorials (stop watching, start doing)

Certificates prove you watched. These prove you can build. This is the layer that actually sticks.

• Interactive Prompt Engineering Tutorial (9 chapters with a live playground): github.com/anthropics/prompt-eng-interactive-tutorial

• Anthropic’s full courses repo (API fundamentals, real-world prompting, evals, tool use): github.com/anthropics/courses

• Official docs: docs.anthropic.com

Step 4: Build ONE real project (the thing that actually gets you in)

This is the step most people will skip. And it’s the only one that decides it.

The application filter is “comfortable working with Claude.” Certificates say you watched videos. A project says you can do the job. So find a local nonprofit (your temple, a food bank, an animal shelter, anything) and automate one annoying thing for them with Claude’s free tier. Their volunteer intake emails. Their donor thank-you notes. Their event FAQ.

One real project beats ten certificates. Put it at the top of your resume with a number attached: “Saved a 3 person food bank 4 hours a week.”

Every link, in one place

Save this list. Send it to a friend.

• Apply: anthropic.com/claude-corps/fellow

• The announcement: anthropic.com/news/claude-corps

• Anthropic Academy (catalog): anthropic.skilljar.com

• Claude Platform 101: anthropic.skilljar.com/claude-platform-101

• Building with the Claude API: anthropic.skilljar.com/claude-with-the-anthropic-api

• Claude Code in Action: anthropic.skilljar.com/claude-code-in-action

• Coursera: AI Fluency foundations: coursera.org/learn/ai-fluency-framework-foundations

• Coursera: full Anthropic catalog: coursera.org/partners/anthropic

• Prompt engineering tutorial: github.com/anthropics/prompt-eng-interactive-tutorial

• Anthropic courses repo: github.com/anthropics/courses

• Docs: docs.anthropic.com

• Claude (free tier, for your project): claude.ai

My honest take

I’ve watched AI headlines scare early career folks for two years now. Every layoff story, every “AI will replace entry level work” prediction.

So when an AI lab spends $150 million specifically on people with the LEAST experience, that’s not a headline to scroll past. That’s a door. And doors like this one have a 5 week timer on them.

If you’re finding value in my AI action letters, please consider subscribing. It keeps me going.

If you know someone early in their career who keeps hitting the experience wall, please share this with them. This program was built for exactly that person.

That’s it from me today.

Till next time. Stay tuned as I will share best resources from both my Harvard and Google networks to bring the best to you. Let’s up skill together. Aspyre higher!

I asked Claude about a business idea, and it immediately agreed to it.

Then, I pitched to it an absolutely unviable idea. I asked it “Should I make rain coats for pigeons”. It told me “YES” again.

Here’s the thing nobody tells you about asking an AI for advice. It will always answer. It will sound sure. And it has no idea what it’s missing, because it only thought about your problem from one angle.

You wouldn’t make a big decision by asking one friend and stopping there. You’d ask the optimist, the cynic, the one who’s done it before, and the one who’ll tell you the ugly truth. Then you’d weigh it.

That’s exactly what Andrej Karpathy built. And someone turned it into a Claude skill you can run in about four minutes.

Let me show you how it works and how to set it up.

What the LLM Council actually is

Karpathy (ex-Tesla AI, OpenAI founding team, the guy whose YouTube lectures taught half of us how transformers work) shipped a little project called llm-council as a Saturday hack.

The idea is simple and kind of brilliant.

Instead of asking one model, you ask several. Then you make them grade each other. Then a final model reads everything and writes the answer.

It runs in three stages:

1. First opinions. Your question goes to every model on the council at the same time. GPT, Gemini, Claude, Grok. Each one answers on its own, blind to the others.

2. Peer review. Now each model gets to see everyone else’s answers, but the names are stripped out. No labels. So a model can’t go easy on a buddy or trash a rival. It just ranks the responses on accuracy and insight, judging the work, not the brand.

3. The Chairman. One designated model reads all the answers and all the rankings, then synthesizes a single final response for you.

The anonymized review in stage two is the clever part. Models have biases about each other. Strip the names and they’re forced to judge the actual reasoning. (Karpathy noted the models are often pretty harsh critics of each other, which is exactly what you want.)

Two ways to run it

There are two flavors of this, and they solve different problems. Pick based on what you actually need.

Option A: Karpathy’s original web app. Real models. GPT-5.1, Gemini 3 Pro, Claude Sonnet 4.5, Grok 4, all answering through your API key. This is the legit multi-model council. Costs money per query (you pay for every model’s tokens), but you get genuine diversity of architectures.

Option B: The Claude skill. Someone rebuilt the concept as a skill that runs entirely inside Claude Code. Instead of four different companies’ models, Claude spawns five “advisors,” each told to think in a totally different style. Same three-stage flow. No second API key, no web server, no setup beyond dropping a file in a folder.

If you live in Claude Code already, start with Option B. It’s the faster win. I’ll cover both.

Setting up the Claude skill (the 4-minute version)

This is the one I’d do first. You need Claude Code installed. That’s it.

Step 1. Drop the skill into your skills folder.

Skills live in a specific spot. One command clones it into place:

git clone https://github.com/tenfoldmarc/llm-council-skill ~/.claude/skills/llm-council

If you’d rather do it by hand: make a folder at ~/.claude/skills/llm-council/, put a SKILL.md file inside it, and restart Claude Code. The folder name and the SKILL.md filename matter. Get those wrong and Claude won’t see it.

Step 2. Restart Claude Code so it picks up the new skill.

Step 3. Trigger it. You don’t type a slash command. You just talk. The skill activates on phrases like:

• “council this”

• “run the council on [your question]”

• “pressure-test this”

• “stress-test this”

• “war room this”

So you’d literally type: council this: should I raise my brand deal rates by 30% going into Q3?

What happens next. The skill scans your workspace context first (it’ll read your CLAUDE.md and memory files if you have them, so it knows who you are). Then it spawns five advisors in parallel, each with a different brain:

• The Contrarian hunts for where this blows up.

• The First Principles Thinker asks if you’re even solving the right problem.

• The Expansionist finds the upside you’re ignoring.

• The Outsider gives you the view with zero context, the dumb-smart question.

• The Executor tells you what to actually do Monday morning.

They peer-review each other anonymously, the Chairman synthesizes, and you get a report showing where they agreed, where they fought, your blind spots, and a recommendation. Takes about four minutes and spits out an HTML report plus a markdown transcript.

The disagreement is the gold. When the Contrarian and the Executor clash, that fault line is usually the exact thing you hadn’t thought hard enough about.

Setting up Karpathy’s original web app

Want the real multi-model version with four different companies arguing? It’s a local web app. A bit more work, maybe 15 minutes.

Step 1. Get the code.

git clone https://github.com/karpathy/llm-council
cd llm-council

Step 2. Install dependencies. Backend is Python (it uses uv), frontend is React.

uv sync
cd frontend && npm install && cd ..

Step 3. Get an OpenRouter key. This is the trick that makes it simple. OpenRouter is one API that routes to every model. So you get one key instead of four separate accounts. Sign up at openrouter.ai, buy some credits, grab your key.

Make a file called .env in the project root:

OPENROUTER_API_KEY=sk-or-v1-your-key-here

Step 4. Pick your council (optional). Open backend/config.py. This is where you choose who sits on the panel:

COUNCIL_MODELS = [
    "openai/gpt-5.1",
    "google/gemini-3-pro-preview",
    "anthropic/claude-sonnet-4.5",
    "x-ai/grok-4",
]
CHAIRMAN_MODEL = "google/gemini-3-pro-preview"

Step 5. Run it.

./start.sh

Then open

http://localhost:5173

in your browser. Ask your question and watch the three stages play out in tabs. Your conversations get saved as JSON files locally, so nothing leaves your machine except the API calls.

Best practices (learned the hard way)

A council is only as good as how you use it. Some no-fluff rules:

1. Only use it for judgment calls, not facts. This is the big one. Don’t council “what’s the capital of France” or “summarize this PDF.” Those have one right answer and a panel just wastes tokens. Use it for pricing, positioning, a pivot, a hire-vs-automate call, messaging that has to land. Anything where reasonable smart people would disagree.

2. Feed it real context. A vague question gets a vague council. Don’t ask “should I raise prices.” Ask “I charge $4K for two LinkedIn posts plus a CEO interview, my engagement is up 40% this quarter, a competitor just quoted a brand $6K, should I raise to $5.5K and what’s the risk.” Specifics in, specifics out.

3. Read the disagreements, not the summary. The Chairman’s final answer is a polished blend. Useful. But the real value is buried in stage two, where the models tore into each other. That’s where the assumption you didn’t question gets exposed.

4. Mix your models if you go the API route. A council of four OpenAI models is just one opinion in four outfits. The whole point is architectural diversity. Keep GPT, Gemini, Claude, and Grok on the panel. Different training, different blind spots.

5. Don’t outsource the decision. The council doesn’t decide. You do. It’s there to surface what you’d have missed, not to hand you a verdict you follow blindly. I’ve overruled the Chairman before and been right.

6. Watch the cost on the API version. Every query hits every model. Four models plus a review pass plus a Chairman is a lot of tokens per question. Great for your ten hardest decisions a month. Not for casual chat.

So which one?

If you want it working in the next five minutes, install the Claude skill and type “council this.” If you’re making a genuinely high-stakes call and want four different companies’ best models in the room, spin up Karpathy’s web app with an OpenRouter key.

Honestly? I use the skill for speed and pull out the full web app when real money’s on the line.

That brand deal I almost mispriced? I ran it through a council. The Contrarian flagged that my anchor number ignored my engagement growth. I went higher. They said yes.

One opinion would’ve cost me real money.

If you are finding value in my AI action letters, please consider subscribing. It keeps me going.

If you know someone who’s about to make a big call with one AI opinion and a prayer, please share this with them.

That’s it from me today.

Till next time. Stay tuned as I will share best resources from both my Harvard and Google networks to bring the best to you. Let’s up skill together. Aspyre higher!

If you’re here, you probably commented “FDE” on my Instagram posts. So first off, thank you. And as promised, this is the whole thing: the complete roadmap, the learning guide, and where to find the actual job openings. All in one place. Bookmark it.

A few weeks ago a friend from my Harvard cohort texted me a screenshot. It was an OpenAI job posting. Total comp north of $500K. The title? “Forward Deployed Engineer.”

His message was three words. “What even is this?”

And honestly, I get the confusion. The name sounds like something out of a military briefing. But this is quietly becoming one of the most valuable jobs in tech right now. While everyone is busy learning AI tools, a small group of people is getting hired to actually deploy AI inside companies.

a16z literally called it “the hottest job in tech.” One report tracked 800% growth in these roles during 2025.

So I went down the rabbit hole. I read the Palantir engineering blog, the Pragmatic Engineer breakdown, comp reports covering 1,200 FDEs, and watched a bunch of day-in-the-life videos. Then I pulled it all into one place.

If you’re an engineer wondering whether to make this pivot, or a student trying to figure out where the puck is going, this is your one-stop guide. Let’s get into it.

(I was recently in NYC; here is a picture taken from one of Google NYC’s iconic offices)

First, what the heck is a Forward Deployed Engineer?

Strip away the buzz and it’s simple.

A Forward Deployed Engineer (FDE) is a software engineer who gets embedded directly inside a customer’s company. Not on a sales call. Not over email. Inside their Slack, inside their infrastructure, sometimes literally sitting in their office for weeks.

They work directly with a business, understand its real problems, and then build AI-powered solutions around its actual workflows. That means connecting AI with the company’s data, internal tools, security systems, automations, and day-to-day operations.

In one line: they’re the bridge between powerful AI models and real-world business problems.

Because here’s the thing nobody tells you. A foundation model on its own doesn’t solve a single business problem. Someone has to wire it into the mess. That someone is the FDE.

Palantir invented this role back in the early 2010s. They called them “Deltas.” Here’s the wild part. Until 2016, Palantir employed more FDEs than regular engineers. The whole company was built on this idea.

The way Palantir frames it stuck with me. A normal developer is “one capability, many customers.” An FDE is “one customer, many capabilities.” You go deep on a single client and solve everything in front of you.

Why this role exploded in 2025 and 2026

Two words. Foundation models.

LLMs are powerful but messy. A company buys access to GPT or Claude and then realizes they have no idea how to wire it into their actual workflows, their actual data, their actual mess. The demo looks magical. The production reality is brutal.

That gap is exactly where FDEs live.

OpenAI formalized the role in early 2025 with Colin Jarvis as Head of FDE. They scaled to 10+ FDEs across 8 cities on 3 continents fast. Anthropic is hiring them. Google, Databricks, Ramp, Gecko Robotics, Commure, Salesforce. The list keeps growing. One job tracker counted 224 open FDE roles across 39 AI companies.

The reason is brutal and simple. AI labs make money when customers actually deploy. And customers don’t deploy on their own. They need someone in the trenches who can write real code on their systems and prove the thing works.

The money (let’s just talk about it)

I’ll be straight with you. This is part of why the role is so hot.

From the 2026 comp data I dug through, here’s the rough landscape:

• Palantir median: around $215K total comp

• Mid-level FDE (market median): about $385K total comp

• OpenAI and Anthropic mid-to-senior: $350K to $550K

• Staff level: $610K and up

• Principal FDEs at frontier labs: clearing $1.2M

A few things to understand before you get stars in your eyes.

Equity is now 55 to 70% of comp at the top of the market. And at OpenAI and Anthropic that equity is tied to private valuations that get revised every six to nine months. So a “$550K” offer in February might be worth a lot more or a lot less by August. It’s a bet, not a salary.

There’s also an FDE premium baked in. These roles pay 15 to 25% more than equivalent ML engineering roles, because the labs benchmark FDEs against research engineers, not against normal solutions architects. You’re being paid like a builder, not a support rep.

What an FDE actually does all day

The romantic version is “you parachute in and save the day.”

The real version is messier and more interesting. Based on how OpenAI structures it, the work runs in three phases.

1. Scoping. A couple of days onsite with the customer. You map their processes, find the real bottleneck, and throw together a quick prototype using synthetic data. Scrappy. Fast. Ugly is fine.

2. Validation. You build evaluation metrics, label data, and stress-test the solution against real benchmarks. This is where you “prove out the brick walls” before anyone commits months to it. Most FDEs say this is the part that separates a real solution from a pretty demo.

3. Delivery. Now you write production code directly on the customer’s infrastructure. Often several days a week onsite. The mantra is “smallest unit possible” that works end to end. Ship something real, then expand.

One FDE described their week as a rough 50/50 split. Half meetings and customer conversations. Half actual building: coding, prompt work, debugging. Expect to travel 20 to 50% of the time depending on the company.

If you hate ambiguity and you only feel safe with a perfectly defined ticket, this role will eat you alive. If you love walking into chaos and turning it into something that works, you’ll love it.

The skills you actually need

I read through about a dozen roadmaps and they all converge on the same four buckets.

1. Strong engineering fundamentals. You need production proficiency in at least one serious language (Java, C++, TypeScript) plus strong Python. You’ll write clean code, build integrations, fix nasty edge cases, and wrangle APIs and data pipelines daily. Problem-solving is the number one skill, full stop.

2. AI and LLM fluency. For the AI roles, you need to actually understand RAG architecture, prompt design, evaluation frameworks, agent failure modes, and the cost-versus-latency tradeoffs. Not the buzzwords. The real mechanics of why an agent breaks and how to fix it.

3. Customer communication. This is the part most engineers underrate. You have to listen to a customer, understand how their business actually makes money, and explain a technical solution in language a non-technical executive gets. Shyam Sankar, Palantir’s President, says the best FDEs are “heretics” and “rebels” with the depth and energy to unlock 3x to 10x growth. Low ego. High empathy. Humble and collaborative.

4. End-to-end ownership. This is the big one. The role basically demands “startup CTO” instincts. You scope it, build it, ship it, and own the outcome.

Here’s a stat that should change how you think about your path. The number one predictor of being a great FDE? Being one of the first 10 engineers at a startup. If you’ve done that, you’ve already lived this job. The ambiguity, the wearing five hats, the shipping under pressure. It’s the same muscle.

How to actually break in

The good news. FDE hiring is portfolio-driven, not credential-driven. Nobody cares about your GPA. They care if you can ship in chaos.

Here’s the path I’d map out if I were doing this today.

Build end-to-end, not toy projects. The most important thing on an FDE resume is evidence that you scoped, built, shipped, and owned something real with measurable impact. One full product beats ten half-finished tutorials.

Get reps with LLMs in production. Build an actual RAG app. Build an agent. Make it break, then figure out why. Write down the failure modes. That hands-on debugging story is gold in interviews.

Practice explaining tech to non-engineers. Record yourself explaining a project to a “customer.” If you can make your mom understand why your solution matters, you’re ahead of most engineers applying.

Use the free training. Salesforce runs a ton of relevant material on Trailhead for free, and their “Ready in Six” onboarding program (technical training, field work, and a capstone) is a great blueprint for what skills to build even before you get hired.

Target the right companies. Early-stage AI startups are the easiest entry point and the best training ground. The frontier labs (OpenAI, Anthropic) are the top of the mountain. Start where you can get reps, then climb.

Where the jobs actually are (the openings I promised)

Companies like OpenAI, Google, Anthropic, and Palantir are all actively hiring for this right now. And they’re not alone. Databricks, Ramp, Gecko Robotics, Commure, and Salesforce are all building out FDE teams. One tracker counted 224 open FDE roles across 39 AI companies.

Here’s exactly where to look:

• Palantir careers (the original, still the biggest): https://www.palantir.com/careers/open-positions/

• OpenAI careers (filter for “Forward Deployed”): https://openai.com/careers/

• Anthropic careers: https://www.anthropic.com/careers

• Levels.fyi to sanity-check comp before you negotiate:

https://www.levels.fyi/

• LinkedIn and Wellfound: search the exact phrase “Forward Deployed Engineer” and set a job alert. New roles drop weekly.

Pro tip. Don’t just apply cold. The early-stage AI startups posting these roles are the easiest to reach. Find the hiring manager on LinkedIn, show them one thing you built, and skip the line.

Resources to go deeper

I’m not going to make you find these yourself. Here’s the actual reading and watching list I’d hand to anyone serious about this.

Read these:

• A Day in the Life of a Palantir FDE (the original source): https://blog.palantir.com/a-day-in-the-life-of-a-palantir-forward-deployed-software-engineer-45ef2de257b1

• The Pragmatic Engineer deep dive (best overall breakdown):

The Pragmatic Engineer

What are Forward Deployed Engineers, and why are they so in demand?

One interesting trend in the AI startup segment this year is the rise to prominence of a specialized software engineering role called “Forward Deployed Engineer” (FDE…

Read more

10 months ago · 403 likes · 6 comments · Gergely Orosz

• PostHog: WTF is a forward deployed engineer? (honest and practical): https://posthog.com/blog/forward-deployed-engineer

• Salesforce: 5 Skills for the FDE role: https://www.salesforce.com/blog/forward-deployed-engineer/

• The New Stack on the FDE hiring race: https://thenewstack.io/forward-deployed-engineer-fde-openai-google/

Watch these:

• Day in the Life of an FDE:

• Day in a Life of a Forward Deployed AI Engineer:

• The New FDE Role Explained (OpenAI, Google, Anthropic hiring):

A copy-paste prompt to map your own gap

Want to know exactly what you’re missing? Drop this into Claude or ChatGPT.

You are a hiring manager for Forward Deployed Engineer roles at a top AI lab.

Here is my background:
[paste your resume or a paragraph about your experience]

Do three things:
1. Score me 1-10 on each of the four FDE skill buckets:
   engineering fundamentals, AI/LLM fluency, customer communication,
   and end-to-end ownership. Be brutally honest.
2. For my two weakest areas, give me a specific 30-day project
   that would close the gap and look great on an FDE resume.
3. Write three interview talking points from my real experience
   that signal "I can ship in ambiguity."

Use it, then go build the projects it spits out. That’s the whole game.

My honest take

The FDE role is what happens when “software engineer” and “the person who actually makes AI work for real companies” become the same job.

It’s not for everyone. If you want a quiet desk and a clean backlog, skip it. But if you like ambiguity, you like talking to humans, and you like the feeling of building something that ships and matters, this might be the best-positioned role in tech for the next five years.

The demand is real. The money is real. And the skills it asks for are the exact skills that survive whatever AI does to the rest of the industry. Because the one thing AI can’t automate is the messy human work of figuring out what a customer actually needs.

This is one of those roles that could define the next decade of tech careers. I genuinely believe that. Get in early.

If you are finding value in my AI action letters, please consider subscribing. It keeps me going.

If you know an engineer who’s been quietly hunting for their next move, please share this with them. It might be the nudge they needed.

That’s it from me today.

Till next time. Stay tuned as I will share best resources from both my Harvard and Google networks to bring the best to you. Let’s up skill together. Aspyre higher!

Disclosure: This article is part of a paid partnership with @Salesforce as part of TDX 2026. All opinions are my own. #ad #SalesforcePartner

Hey folks,

I just got back from 2 days at Salesforce TrailblazerDX in San Francisco, and I’m still processing.

I went in expecting product launches. New features. The usual.

What I left with was bigger than that.

It was a question I can’t shake.

If you can build a real, deployed AI agent on a live website in under 10 minutes (no Salesforce experience, no setup, just a laptop and a prompt), what does the next 5 years of work actually look like?

I watched it happen on stage. Live build. Live deploy. They found a bug, fixed it from Slack, and pushed the update. Then they shipped that same agent to 6 surfaces. Website, Slack, Teams, ChatGPT, Claude, mobile. Zero rebuilds.

The conversation that pulled it all together

Between sessions, I sat down with Sally Al Ghoul. She’s 15x Salesforce certified, founder of Code with Sally (she teaches Salesforce in English and Arabic to thousands of developers across the Middle East), and co-founder of Trailblazing Teens.

I asked her one question:

What % of code in 5 years will be written by AI?

Her answer: 90 to 100%.

But here’s the part I keep replaying:

“The role of the developer doesn’t disappear. It shifts. Less typing. More designing, reviewing, orchestrating the agents.”

Then I asked her what’s actually blocking enterprise AI right now. Her answer was the same thing the 55% stat told me:

“It’s mainly the human being. Being able to adapt to the new mindset shift. AI is changing super quick. We have to keep learning, adapting, experimenting.”

Same theme. Different angle. The bottleneck is us.

The full interview is up on YouTube. If you’re a Salesforce dev, a builder, or a PM trying to figure out where this is heading, watch the whole thing.

I also interviewed a SalesForce Expert Sally El Ghoul (15x Salesforce Certified Developer):

That used to be months of work across 4 teams.

It happened in 10 minutes.

So this newsletter is my attempt to share what I saw, what I felt, and the one stat that genuinely stopped me in my tracks.

Let’s go.

The stat that stopped me

55% of enterprise companies say their biggest AI problem isn’t tools.

It’s talent.

Read that again. They want agents. Their boards want agents. They just don’t have enough people who can build them.

For the last 2 years we’ve been told the bottleneck to AI was the technology. It wasn’t. It was always the builders. And TDX 2026 made that brutally clear.

If you’re upskilling in AI right now, you’re not late. You’re early.

The 6 things I got hands on with

I sat down at the TDX campground and built a working agent. No Salesforce background. Just me, a laptop, and Agentforce Vibes 2.0. Here’s what I tested:

1. Slackbot joining Google Meet. Not just transcription. It reads your prior Slack context. So if you join 15 minutes late, it catches you up in the actual context of conversations you were already having.

2. Slackbot department skills. One bot, six tabs replaced. Competitive intel, CRM lookups, web search, past chat history, draft and send emails, marketing tool integration. All in Slack. No more switching.

3. Agentforce Labs. They let you sit down and build. Not watch a demo. Free Developer Edition with Claude Sonnet access included.

4. Custom skills on private infra. Orgs can now use Claude Code to build their own Agentforce skills on their own private environments. Your data stays yours.

5. Tableau dashboards read by agents. Ask an agent a question about your data. It pulls the right Tableau dashboard, reads it, gives you the answer. No more digging through 30 dashboards.

6. Open Telemetry for agents. Every action your agent takes is traceable. What it did, why it did it, what data it pulled.

The numbers that hit hardest

A few stats from the keynotes:

• 100M lines of agentic code already accepted through Agentforce Vibes

• 25% of developer code is now AI-generated (and that number is rising fast)

• 22x surge in Agentforce Vibes usage since launch

• Lumen: 300+ hours per week saved on contract renewals

• ADP: AI handling HR for the company that processes 1 in 6 US paychecks

• PenFed: 75% of routine IT tasks now handled by agents

These aren’t pilots. They’re live in production right now.

Two new job categories nobody is hiring for yet

Here’s where I think the real opportunity is. After 2 days at TDX, I’d bet on two roles emerging fast:

Skill engineering. Building reusable agent capabilities (a CRM lookup skill, a contract-review skill, a customer-support skill) that orgs can plug in across their stack. Right now, most companies are reinventing these from scratch. The teams who package them well will be in massive demand.

UI engineering for agents. Agents don’t render their own interfaces. They pull context from MCP servers and generate UI dynamically. The people who design those generative interfaces (text, voice, vision, all working together) are going to be the new front-end engineers. Just for an agent-first world.

If you’re early in your career, these are the two areas I’d be watching like a hawk.

My honest take after 2 days

Parker Harris asked one question on stage that summed up the whole event:

“Why should you ever log into Salesforce again?”

It sounds like marketing. It’s not.

The platform is going agent-first. Every service is now an API, an MCP tool, or a CLI command (they’re calling this Headless 360). Agents call APIs, invoke tools, run commands. They don’t click through menus.

If you’re a builder, your job stops being “click through the UI” and starts being “design the workflow the agents will run.”

If you’re a PM, your job stops being “ship the next feature” and starts being “decide what outcomes the agents should pursue.”

If you’re an exec, your job stops being “approve the headcount plan” and starts being “decide which workflows are agent-ready and which still need humans.”

The companies that figure this out won’t just be faster. They’ll be running a different business entirely.

What I’m doing about it

I’m going to spend the next month doing 3 things:

1. Building a skill in Agentforce Vibes 2.0. Just to prove I can. I want to feel the workflow myself.

2. Recording more conversations like the one with Sally. The Shortcast lineup for the next quarter is going to lean into builders who actually have agents in production.

3. Sharing what I learn here in this newsletter. Every Sunday. No fluff. Just what’s working and what’s not.

If you are finding value in my AI action letters, please consider subscribing. It keeps me going.

If you know someone who’s a Salesforce developer, an ops leader, or a PM trying to make sense of agentic AI, please share this with them. This is the kind of moment they shouldn’t miss.

That’s it from me today.

Till next time. Stay tuned as I will share best resources from both my Harvard and Google networks to bring the best to you. Let’s up skill together. Aspyre higher!

Abhijay

#SalesforcePartner #TDX2026 #Agentforce #AI #ad

Two weeks ago I posted a carousel on LinkedIn about Claude certifications. 2,400 saves. That’s not likes. Not comments. Saves.

People are hoarding this information like it’s a coupon code that’s about to expire.

And honestly? It kind of is. Because right now, the Claude certification market is in that tiny window where the certs exist, they’re free, and almost nobody has them on their profile yet. That window closes fast. It happened with Google’s AI certs in 2023. It happened with AWS in 2019. Early movers got interview callbacks. Everyone else got “oh cool, another cert.”

So here’s the play. Anthropic quietly launched Anthropic Academy. 17 courses total. No credit card. No API key required. Just your email.

But you don’t need all 17. You need 13.

I went through every single one and picked the 13 that actually move the needle on a resume, a LinkedIn profile, or a real project. The rest are audience-specific variants you can skip unless they match you exactly.

Let’s get into it.

The 13 Certs (Grouped by Track)

Track 1: AI Fluency (Start Here)

These are for everyone. PM, marketer, founder, student, engineer. Doesn’t matter.

1. Claude 101

The starting point. Core features, prompting basics, real use cases. About 2 hours. If you’ve been using Claude daily, you’ll breeze through it. Take it anyway. It fills gaps you didn’t know you had.

2. AI Fluency: Framework and Foundations

This one changed how I think about AI. It teaches Anthropic’s 4D framework: Delegation, Description, Discernment, Diligence. 14 lessons. It’s not about Claude specifically. It’s about building the muscle of knowing what to hand off to AI and what to keep. I use this framework every day at Google.

3. AI Capabilities and Limitations

New as of April 2026. This is the “honest” course. What LLMs can actually do, where they break, why they hallucinate. If you’re going to put Claude certs on your resume, you need to be able to talk about limitations in an interview. This is how you prep for that.

Track 2: Builder (The Meaty One)

This track is where the real differentiation happens. The majority of people stop after the fluency courses. Don’t be the majority.

4. Claude Code 101

Claude Code is Anthropic’s coding agent. It reads your files, runs commands, edits code, commits to Git. This course covers the Explore, Plan, Code, Commit workflow. If you’ve never used Claude Code, start here before the next one.

5. Claude Code in Action

The advanced version. Context management, custom workflows, hooks, integrating with external services. This is the course I recommend most to engineers. It’s also available on Coursera if you want it on that profile too.

6. Building with the Claude API

The biggest course in the catalog. Auth, streaming, multi-turn conversations, structured outputs, tool use, building AI agents. Even if you think you know APIs, take this one. The section on tool use alone is worth the time.

7. Intro to Model Context Protocol (MCP)

MCP is how Claude connects to external data. Your calendar, your CRM, your codebase, your database. This course teaches you to build MCP servers from scratch in Python. It covers the three core primitives: tools, resources, and prompts. You can also check out the official MCP spec to go deeper.

If there’s one cert on this list that will age the best, it’s this one. MCP is becoming the USB-C of AI integrations.

8. MCP: Advanced Topics

Production patterns. Transport mechanisms, deployment, reliability, debugging. Take this right after the intro course while the concepts are fresh.

9. Introduction to Agent Skills

Skills are how Claude reads markdown files to execute repeatable workflows. Think of them as custom commands you teach Claude. This is the newest concept, and honestly the easiest one to demo on LinkedIn. Build a skill, screen record it, post it. I’ve seen people get 500K+ impressions doing exactly that.

10. Introduction to Subagents

Added April 2026. This teaches you how to delegate isolated tasks to sub-agents within Claude Code. Think of it like being a tech lead who assigns tickets to junior devs, except the junior devs are AI agents running in parallel.

Track 3: Cloud and Enterprise

These two are for people deploying Claude in production at companies. If you’re a student or solo builder, you can skip these for now and come back later.

11. Claude with Amazon Bedrock

How to deploy and use Claude natively on AWS. API calls, tool use, retrieval patterns, all through Bedrock.

12. Claude with Google Cloud Vertex AI

Same thing but on GCP. Streaming, tool use, deployment via Vertex. (I’m biased because I work at Google, but if your company is on GCP, this is the one to take.)

Track 4: The Wildcard

13. Introduction to Claude Cowork

Also added April 2026. Cowork is Anthropic’s agentic desktop assistant. It’s different from Claude Code. Instead of coding tasks, you delegate knowledge work. Research, analysis, writing, spreadsheet manipulation. Think of it as Claude that can use your computer.

This cert is a sleeper. Most people haven’t heard of Cowork yet. That’s exactly why you want it on your profile now.

The 90-Day Plan (Don’t Just Collect Them)

Here’s my honest take. Certs without projects are wallpaper. Projects without certs are invisible to recruiters. You need both.

Week 1 to 2: Claude 101, AI Fluency Framework, AI Capabilities and Limitations. Don’t skip the exercises.

Week 3 to 4: Claude Code 101 and Claude Code in Action. Pick a real project while you do these. A resume parser, a meeting summarizer, a Slack bot. Something you’d actually use.

Week 5 to 6: Building with the Claude API. Migrate your project from Claude Code to a proper API integration.

Week 7 to 8: Both MCP courses back to back. Build one MCP server that connects your project to a real data source. Your inbox, your Notion, your CRM.

Week 9: Agent Skills and Subagents. Wrap your project’s repeatable workflows as skills.

Week 10: Claude Cowork. Then pick one cloud course based on whatever cloud your company uses.

Week 11 to 12: Go back and re-take the quizzes for any course where you scored below 90%. Make sure the concepts stuck.

The key is the project running parallel to the certs. You finish week 12 with 13 certifications AND a shipped project. That’s a story. That gets interviews.

How to Put Them on LinkedIn (Most People Mess This Up)

When you finish a course, Skilljar emails you a certificate with a unique verification URL. Don’t just screenshot it and post.

Go to your LinkedIn profile. Click “Add profile section.” Click “Licenses and certifications.” Fill it out:

• Name: exact course title from Skilljar

• Issuing organization: Anthropic

• Issue date: whatever the cert says

• Credential URL: paste the Skilljar verification link

That way recruiters can click through and verify it’s real. Most people just upload a JPEG. Don’t be most people.

Stack them with the most impressive first. For most roles, that’s Building with the Claude API, MCP Advanced, and Claude Code in Action.

Quick Tips That’ll Save You 10 Hours

A few things I wish someone had told me.

The video player supports 1.25x and 1.5x speed. Use it. The content is great but the pacing is slow.

Take the quiz before watching the lecture. Bomb it on purpose. Then watch the video. Then re-take. You’ll retain way more.

Don’t take notes. Build instead. The Anthropic Cookbook on GitHub has working Jupyter notebooks for every concept the courses cover. Clone it, run it, break it.

Download the PDF certificates the day you earn them. Back them up to Google Drive. Skilljar dashboards get weird after long periods of inactivity.

Before you start the developer track, grab a free API key at console.anthropic.com (new accounts get $5 free credits) and run through Anthropic’s prompt engineering interactive tutorial on GitHub. It’s 9 chapters of hands-on exercises and it’ll make the API course click way faster.

Bonus: The Paid One Worth Knowing About

After you’ve done all 13, there’s one more. The Claude Certified Architect Foundations exam. $99 per attempt. 60 scenario-driven questions, proctored, 120 minutes. It covers agentic architecture, Claude Code workflows, prompt engineering, tool orchestration, and context management.

You don’t need this to get hired. But if you’re gunning for enterprise AI consulting roles or selling Claude solutions to companies, it’s the closest thing to an “AWS Solutions Architect” for the Claude ecosystem. Worth the $99 if that’s your path.

If you are finding value in my AI action letters, please consider subscribing. It keeps me going.

If you know someone who’s been applying to jobs without any AI credentials on their resume, please share this with them. These 13 certs are the single best free credential play of 2026.

That’s it from me today.

Till next time. Stay tuned as I will share best resources from both my Harvard and Google networks to bring the best to you. Let’s up skill together. Aspyre higher!

Two weeks ago, I did something I should have done a year ago.

I uploaded my own LinkedIn profile to Claude and asked it to tear me apart.

Same profile that pulls about 3,000 views a week. Google PM. Harvard MBA. 77K followers. I figured Claude would give me a polite pat on the head, point out a typo, and move on.

It didn’t.

My old headline buried the most interesting thing I do. My About section opened with this gem: “I’m a Product Manager at Google passionate about building products that delight users.” I’m cringing typing it. Possibly the most generic sentence ever written by a human about themselves. My experience bullets were verbs and vibes. Zero numbers. Nothing a recruiter could actually evaluate.

So I rebuilt the whole thing Tuesday night and pushed it live before bed.

By Friday, I had 4 inbound recruiter DMs. Two from companies I’d never heard of. Two from places I’d be open to discussing. ( I am not actively job hunting, YouTube is great!)

I didn’t apply to anything. The profile did the work.

For reference, here’s an example of what the new opening line could look like now:

I help YouTube ship the AI features that 2.7B people use every month. Previously the team behind viewer experience for the YouTube app on mobile.

Specific. Scoped. Numbers. No “passionate.” That single swap is what most of the work below is really about.

This is the exact 5-prompt workflow I used. I’m sharing it because most of the LinkedIn-with-AI guides out there hand you one massive prompt and call it a day. That’s exactly why their rewrites still read like ChatGPT poetry.

Why one mega-prompt doesn’t work (and what to do instead)

Every LinkedIn coach has a 2,000-word prompt that asks Claude to “audit your headline, photo, About, experience, skills, and featured section in one go.”

Don’t do that.

Claude’s output quality drops fast when you stack 6 jobs into one ask. You end up with shallow rewrites for every section instead of a brilliant rewrite of the one section that actually matters. Sequence beats stacking.

Here’s the other thing every AI-LinkedIn guide does wrong: they tell you to enable the Apify connector so Claude can scrape your live profile. Skip that. Apify is for high-volume scraping of other people’s data. For your own profile, it’s a flaky middle layer that breaks on private profiles, mobile sessions, and rate limits. Half my readers who tried it last month hit one of those failure modes.

The reliable path is dumber and works every time: export your LinkedIn as a PDF and upload that.

The 3 reasons most AI LinkedIn rewrites flop

1. No target role. Claude can’t optimize you for “a better job.” It needs the actual JD of what you’re applying to. Without that, you get corporate prose that could belong to anyone with a pulse.

2. No real numbers. “Led cross-functional initiatives” tells a recruiter nothing. “Shipped a YouTube Studio feature now used by 4M+ creators monthly” tells them everything. Claude will not invent your metrics. You have to feed them in.

3. One-shot prompting. Asking Claude to rewrite your whole profile in one go produces the same generic LinkedIn-ghostwriter voice for every section. Sequencing fixes this.

Setup (under 3 minutes)

Two files. That’s it.

• Your LinkedIn as a PDF. Open your profile, click “Resources” on the right side, click “Save to PDF.” Download.

• 3 to 5 job descriptions for roles you actually want. Copy the full JD text into a single Google Doc.

If you’re chasing roles that barely exist yet (AI PM, AI ops, agentic workflows lead), grab the closest 5 JDs you can find. Even imperfect targets give Claude something concrete to anchor on. Vague targeting is what produces vague output.

Open claude.ai. Start a new chat. Drop both files in before you type anything.

Now run the 5 prompts in order.

Prompt 1: The brutal audit (do not skip this)

Most people jump straight to rewriting. That’s why their rewrite still misses. You need a diagnostic before you start cutting.

You are my friend who has been a senior tech recruiter at [target companies like Stripe, OpenAI, Anthropic] for 10 years. You owe me honesty, not politeness. I'd rather hear this from you than from a hiring manager who silently passes.

I'm attaching my LinkedIn profile (PDF) and 5 job descriptions I want to be competitive for.

Do NOT rewrite anything yet. Give me a diagnostic:

1. Identity gap: In one sentence, what identity does my profile currently project? In one sentence, what identity are these JDs hiring for? Where is the mismatch?

2. Keyword gap: Which exact terms appear in 3+ of the JDs that my profile does not mention? List them.

3. Credibility gap: What proof points do these JDs care about (specific tools, scale, metrics, named launches) that my profile buries or omits?

4. Search visibility: If a recruiter typed the 3 most likely LinkedIn search queries for these roles, would my profile show up in the top 10 results? Why or why not?

5. The one line: If my profile had to win on a single sentence, what is it saying today vs what it should be saying?

Quote my actual profile lines back to me. Do not soften.

The “friend who owes me honesty” framing changes everything. Claude’s default with the standard “act as a recruiter” prompt is corporate politeness. With this framing it actually marks up your profile.

Prompt 2: Headline and About rewrite (with hard constraints)

Now that Claude knows what’s broken, give it the rebuild constraints.

Based on the audit above, rewrite my headline and About section.

Headline rules:
- Max 120 characters
- Must include the exact role keyword from at least 3 of the attached JDs
- Lead with the most credibility-dense fact I have (company name, scale of impact, named launch)
- Banned words: passionate, results-driven, transformational, dynamic, seasoned, strategic
- Give me 3 versions, ranked, with one sentence explaining why each one works

About rules:
- Max 2,000 characters
- First line is a HOOK, not "I'm a [title]"
- Paragraph 1: What I build and the scale (with numbers)
- Paragraph 2: 3-sentence origin story (only if it actually differentiates me)
- Paragraph 3: What I'm working on now and what conversations I want to have
- End with ONE specific CTA (not "let's connect")
- Mobile-first formatting: line breaks every 1-2 sentences, no walls of text
- Give me 2 full versions so I can pick

The mobile-first formatting note is what most prompts miss. 80% of LinkedIn views happen on mobile. Wall-of-text About sections get scrolled past in 0.4 seconds. The line breaks are not aesthetic. They’re functional.

Prompt 3: Experience bullet surgery (the part recruiters actually read)

Bullets are where recruiters decide if you can do the job. Most people leave them as the auto-populated job-description filler LinkedIn drops in.

For each role in my Experience section, rewrite the bullets using this exact structure:

[Strong verb] + [specific scope or scale] + [measurable outcome] + [business impact]

Rules:
- 3 bullets per role MAX
- Every bullet must contain at least one number (% change, $ amount, user count, time saved, team size). If you do not have my number for a bullet, flag it as [METRIC NEEDED] so I can fill it in.
- Lead with the most impressive bullet, not the most chronological one
- Use the EXACT vocabulary from the 5 attached JDs. If the JDs say "agentic workflows," do NOT write "AI assistants."
- Banned verbs: spearheaded, championed, drove alignment, leveraged synergies, owned the vision, delivered on

For each role, show me:
- Current bullets (so I see the before/after side by side)
- Rewritten bullets
- A list of [METRIC NEEDED] gaps I have to fill in myself

The [METRIC NEEDED] flagging is the unlock here. Claude won’t make up numbers. It tells you exactly where you need to dig up data from old performance reviews, dashboards, or Slack threads. That’s the work everyone else skips and why their bullets still feel vague after a rewrite.

Prompt 4: Featured section + Skills strategy

Most people ignore Featured. Mistake. It sits right under your photo and gets clicked more than any single bullet in your experience.

Based on my target JDs and my rewritten profile, design my Featured section and Skills section.

Featured (4 slots max):
- What goes in each slot, in what order, optimized for recruiter scan
- For any slot where I don't have content yet, suggest the SPECIFIC piece I should create (a post, a one-page case study, a portfolio link, a Loom)
- For each slot, give me the one-line description that maximizes click-through on mobile

Skills:
- Top 10 skills to pin, ranked by relevance to my target JDs
- Identify which 3 should go in the "Top Skills" slot at the top of my profile (LinkedIn's algorithm uses this to surface me in recruiter searches)
- Call out any skill that appears in 3+ JDs but is MISSING from my current profile
- Suggest 5 people from my network I should ask for skill endorsements, plus the exact 2-line DM template to send each one

Skills is not a checklist. It’s a signal to LinkedIn’s recommender about which roles to surface you for. Getting your top 3 pinned correctly moves the needle more than any other section on the profile.

Prompt 5: The 4-week authority stack

A polished profile with a dead activity feed reads as inauthentic. Recruiters check your posts. If the most recent thing is from 2022 it doesn’t matter how good your About is.

Based on my new positioning, generate a 4-week LinkedIn content plan to build authority BEFORE I start any cold outreach.

Structure:
- Week 1: 3 builder credibility posts (something I shipped, learned, or built)
- Week 2: 3 lessons-from-current-role posts (specific, slightly contrarian)
- Week 3: 3 POV posts on my domain (a take I would defend in an interview)
- Week 4: 3 behind-the-scenes posts (what my actual day-to-day looks like)

For each of the 12 posts, give me:
- Hook (first line, must work on mobile, must stop a scroll)
- Body angle (3-5 lines of structure, not the full post)
- Format (text post, carousel, document upload, short video)
- The ONE specific question to end with to drive comments

Sequence them so they compound. By the end of week 4, a recruiter checking my activity feed should see a coherent narrative about who I am and what I build.

This is the prompt I get the most pushback on. People want to fix the profile and stop. But the dead feed kills the rewrite. Spend 30 minutes a week posting one of Claude’s drafts and the compounding shows up by week 3.

Pro move: turn this into a reusable Claude Project

Do not run this as a one-off chat that you lose in your history.

Open Claude, create a new Project, drop into the Project knowledge:

• Your LinkedIn PDF

• Your 5 target JDs

• A short voice doc with 3-5 samples of your best-performing posts and one or two real emails you’ve written, so Claude learns how you actually talk

Now every new chat in that Project starts grounded in your data and your voice. Three months from now when you ship something new and need to update your About, you don’t re-upload anything. Just open the Project and say “Update my About to reflect the new launch.” Done in 4 minutes.

I keep my own LinkedIn as a Claude Project and refresh it quarterly. 20 minutes per quarter. The compounding effect on inbound is what made me 4-for-Friday two weeks ago.

(If you missed last week’s AI Action Letter, the Claude job hunting carousel I posted on LinkedIn pairs perfectly with this. That covers job scraping, resume rewrites, and warm-intro mapping. This piece is the LinkedIn-profile half of the same stack.)

The 24-hour rule

The biggest mistake I see: people run all 5 prompts, get a beautiful rewrite, and then sit on it for 3 weeks “polishing.”

Push it live within 24 hours.

You can iterate on a live profile. You cannot iterate on a draft sitting in your Claude history. The people who win at this push the rewrite live, watch their analytics for 2 weeks, and refine from real signal. The people who lose keep tweaking the About section while their broken profile stays up and recruiters keep skipping past them.

Done beats perfect. Especially on LinkedIn.

P.S. on the visa side of all this

A bunch of you DM me every week about F1, OPT, STEM OPT, and H1B timing. I’m on STEM OPT myself, so I get the underlying anxiety. The “I have to find a sponsor before X date” math is the loudest thing in your head when you’re job hunting on a visa.

I’m not a lawyer. But I work closely with the team at Manifest Law and they’re who I send everyone who asks. Most immigration lawyers don’t actually understand what it’s like to be on a visa and trying to job hunt at the same time. The Manifest team does.

If you have questions around:

• F1, OPT, STEM OPT timing

• H1B, cap-gap, or employer sponsorship

• Job change, layoff, or status planning

• Long-term green card pathways

They do a free 15-minute consultation through my partner link: https://manifestlaw.com/partner/abhi

If you’re finding value in my AI action letters, please consider subscribing. It keeps me going.

If you know someone who is rewriting their LinkedIn this month, please share this with them.

That’s it from me today. Share with a friend who is job hunting in the US!

Till next time. Stay tuned as I will share best resources from both my Harvard and Google networks to bring the best to you. Let’s up skill together. Aspyre higher!

Last Thursday my friend DMed me at 2am his time.

He’s a backend engineer in Hyderabad, four years of experience, and just got laid off in the second wave of cuts at his startup.

He’d booked the Claude Certified Architect exam for the following Monday and was panicking.

I sent him a folder. Inside it were all the resources I’d collected from friends. One of them was from HBS! Yes, an MBA grad also has passed this certification!

Since then, I’ve gotten the same question maybe forty times.

So I figured I’d write the whole list down. Every single thing. The 13 Anthropic certifications, the official prep stack, the practice question approach, the GitHub repos, and the n8n automation. All of it. With links.

This is the longest issue I’ve ever written. Bookmark it. You’ll come back.

Why this certification actually matters

The Claude Certified Architect is currently the highest-paid AI cert in the market. Recruiters at Snowflake, Databricks, Salesforce, and Anthropic itself screen for it. The passing score is 720 out of 1000. Proctored. Real bar.

It’s also one of the only AI certifications where the test is genuine. You can’t memorize your way through it. You need to actually understand context windows, agentic patterns, MCP, tool use, prompt caching, and how to architect a multi-agent system.

So, let’s get into the actual resources.

Resource 1: Anthropic Academy on Skilljar (the official source)

Start here. The official Anthropic learning portal is at anthropic.skilljar.com. As of April 2026, they have 18 courses across 5 tracks, all free, all with official Anthropic-branded certificates you can post to LinkedIn.

Sign up with just an email. No credit card. No API account required to start.

The course catalog includes:

• Claude 101: foundations of working with Claude

• Claude Code 101: Introduction to the Claude Code CLI

• Building with the Claude API: the bread and butter for engineers

• Introduction to Model Context Protocol: MCP fundamentals

• Advanced MCP Development: building MCP servers and clients in Python

• Introduction to Claude Cowork: Claude for knowledge work

• Introduction to Subagents: multi-agent patterns

• AI Capabilities and Limitations: what Claude can and can’t do

• AI Fluency for Educators, Students, Nonprofits, Small Businesses: role-based tracks

• Claude on Amazon Bedrock: enterprise deployment on AWS

• Claude on Google Vertex AI: enterprise deployment on GCP

If you complete the relevant courses, you have roughly 80% of what the Architect exam tests.

Resource 2: Anthropic’s official courses GitHub repo

This is the public equivalent of the “official prep guide” everyone asks about. It’s not a single 40-page PDF. It’s something better.

github.com/anthropics/courses

The repo contains the full Anthropic curriculum as interactive Jupyter notebooks. It includes:

• API Fundamentals

• Prompt Engineering Interactive Tutorial (9 chapters, the most thorough prompt engineering material Anthropic publishes)

• Real World Prompting

• Prompt Evaluations

• Tool Use

You can clone it, run it locally, and work through every concept the Architect exam tests. This is the closest thing to the rumored “40-page Anthropic guide.” Compiled, it’s well over 40 pages of material.

Resource 3: The Prompt Engineering Interactive Tutorial

Separate, deeper version of the prompt engineering material:

github.com/anthropics/prompt-eng-interactive-tutorial

Nine chapters with exercises plus an advanced appendix. Available as Jupyter notebooks for engineers and as Google Sheets for non-developers who want to use the Claude for Sheets extension.

Resource 4: The Anthropic Cookbook

The single most useful repo if you’re building things with Claude:

github.com/anthropics/anthropic-cookbook

Official Anthropic recipes for every pattern in the API. Prompt caching, citations, batch processing, vision, PDF parsing, tool use, agentic loops, RAG. The exam asks system-design questions that map almost one-to-one to these notebooks.

Pick three notebooks, run them locally, and modify them. That single exercise teaches more than ten hours of video.

Resource 5: The Claude Agent SDK

github.com/anthropics/claude-agent-sdk-python

The official SDK for building custom agents that wrap Claude with tools, hooks, and subagents. The “agents” section of the exam is built around concepts from this SDK. If you’ve built one custom agent end-to-end, you’ll never miss a question in this section.

Resource 6: Claude Code (the CLI)

github.com/anthropics/claude-code

The Claude Code section of the Architect exam tests this stuff directly. Reading is fine. Actually using it for a week before the exam is what makes the difference.

Things to know cold for the exam:

• CLAUDE.md files and how they’re loaded

• Slash commands and skills

• Hooks (Stop, PreToolUse, PostToolUse, etc.)

• The subagent system

• MCP server config

• Permission modes

Resource 7: Model Context Protocol (the spec itself)

modelcontextprotocol.io

MCP shows up everywhere on the exam now. The spec site has the protocol details. Pair it with the server library:

github.com/modelcontextprotocol/servers

This community-maintained repo lists every MCP server, with reference implementations in TypeScript and Python. Clone one. Modify it. Run it against Claude Code. Done.

Resource 8: The Anthropic API console

console.anthropic.com

You’ll need a free account for the hands-on sections of the courses. Different from your Skilljar account. The console also has the Workbench, which is where you’ll do most of your prompt experimentation.

The Workbench includes prompt caching analytics, tool use visualization, and the prompt evaluator. All of these appear in exam questions.

Resource 9: Anthropic’s full documentation

docs.anthropic.com

The single source of truth for every API capability. Three sections to read cover to cover before the exam:

• Build with Claude (covers prompt engineering, tool use, vision, citations, files, PDF support)

• Agents and tools (parallel tool calling, the agentic loop, computer use)

• Test and evaluate (prompt evaluations, the Workbench evaluator)

If you only read three things from this list, make them these.

Resource 10: The Anthropic Learn portal

anthropic.com/learn

The umbrella landing page that links out to Academy, docs, the cookbook, and the research blog. Useful as a single bookmark to keep open during prep.

Resource 11: n8n for the agentic workflow questions

n8n.io

n8n shows up more and more on the exam, especially in the “agentic workflows” section. Build one workflow that uses Claude as the LLM node and routes tool calls. A simple Gmail-to-Slack triage workflow with classification and reply drafting will teach you everything the exam tests on this topic.

If you can explain how the workflow uses tool calling, parallel execution, and prompt caching, you’ll handle every agentic-pattern question on the exam.

Resource 12: The pricing and model selection pages

docs.anthropic.com/en/docs/about-claude/pricing

There’s an exam section that asks you to design cost-optimized systems. You need the current pricing for every Claude model, every tier, and the prompt caching discount tiers locked in your head.

Spend 30 minutes with this page the day before the exam.

Resource 13: Claude Code best practices on the docs site,

docs.anthropic.com/en/docs/claude-code

The full Claude Code docs cover hooks, MCP, slash commands, skills, and subagents. Read the whole tree. The Architect exam reuses these terms verbatim in its questions.

How to actually use this list (the 7-day plan)

If you have a week before the exam:

• Day 1: Skim the docs.anthropic.com “Build with Claude” section. No notes yet, just absorb.

• Day 2: Sign up at anthropic.skilljar.com and knock out Claude 101 + Building with the Claude API.

• Day 3: Clone anthropic-cookbook. Run three notebooks. Modify one.

• Day 4: Work through the prompt-eng-interactive-tutorial end to end.

• Day 5: Clone an MCP server from modelcontextprotocol/servers, run it, and modify it. Build a simple n8n workflow.

• Day 6: Complete the MCP courses and the Subagents course on Academy.

• Day 7: Re-read docs.anthropic.com with notes. Lock in the pricing page. Sleep early. Eat protein in the morning.

If you have a month, stretch it out. If you have a weekend, focus on Days 1, 2, 3, 7. Skip the hands-on if you absolutely must, though it’ll cost you in the system-design section.

Quick reference: every link in this article

• anthropic.skilljar.com: Anthropic Academy (the 18-course catalog)

• anthropic.com/learn: Anthropic Learn umbrella page

• docs.anthropic.com: official docs

• console.anthropic.com: API console and Workbench

• github.com/anthropics/courses: official courses repo

• github.com/anthropics/prompt-eng-interactive-tutorial: 9-chapter prompt engineering tutorial

• github.com/anthropics/anthropic-cookbook: official recipes

• github.com/anthropics/claude-agent-sdk-python: agent SDK

• github.com/anthropics/claude-code: Claude Code CLI

• modelcontextprotocol.io: MCP spec

• github.com/modelcontextprotocol/servers: MCP server library

• n8n.io: workflow automation tool

That’s it from me today.

Till next time, stay tuned as I will share the best resources from both my Harvard and Google networks to bring the best to you. Let’s upskill together. Aspyre higher!

Last month I opened my Anthropic billing dashboard and almost spit out my coffee.

$1,847 in one week. For a side project I built in two evenings.

I went digging. Turns out I was sending the same 4,000-token system prompt on every single API call. Then asking the model to “explain in detail.” Then dumping the full chat history into every follow-up. Classic rookie stuff.

I cut it down to $312 the next week. Same app. Same quality. Just smarter token usage.

If you’re building anything on top of LLMs (a chatbot, an agent, a RAG app, even a Cursor or Claude Code workflow), this stuff matters. Tokens are the new compute bill. And most people are paying way more than they should.

Here’s what actually moved the needle for me. Real numbers, real examples, no fluff.

Why this matters before we get into the how

Two facts to keep in mind for everything below:

1. Output tokens cost 3 to 5x more than input tokens. On Claude Sonnet 4.6, input is $3/M and output is $15/M. So cutting one output token saves you what cutting four input tokens would.

2. Most teams waste 40 to 60% of their token budget on dumb stuff. Repeated system prompts. Bloated context. Models that ramble. Wrong model for the job.

Fix those two and you cut your bill in half before doing anything clever.

Now the tactics.

1. Turn on prompt caching (this is the big one)

If you only do one thing, do this.

Prompt caching lets you mark parts of your prompt as cacheable. The provider stores it. Next time you send the same chunk, you pay 10% of the input price instead of full rate. Cache writes cost 25% more upfront, but if you’re hitting the same system prompt over and over, you break even after about two calls.

I had a research agent calling Claude with a 6,000-token system prompt every time. Once I added cache breakpoints, my input cost dropped 87%.

In Anthropic’s API it looks like this:

client.messages.create(
    model="claude-sonnet-4-6",
    system=[
        {
            "type": "text",
            "text": "You are a research analyst...",  # long stable prompt
            "cache_control": {"type": "ephemeral"}
        }
    ],
    messages=[{"role": "user", "content": user_query}]
)

OpenAI does this automatically now if your prompt prefix is over 1,024 tokens. So the order matters. Put the static stuff (system prompt, examples, knowledge base) at the top. Variable stuff (user query) at the bottom.

Order is everything for caching.

2. Tell the model to be brief. Then enforce it.

Asking for shorter answers is the highest-ROI prompt change you can make. Remember, output is 3 to 5x more expensive than input.

I went from “Explain this concept” to “Explain this in 3 bullets, max 15 words each.” My average response went from 380 tokens to 62 tokens. That’s an 83% cut on the expensive side of the bill.

Even better: use max_tokens as a hard cap. If you only need 200 tokens of output, set max_tokens=200. The model won’t write a novel if it physically can’t.

A trick I use:

Respond in the format below. No preamble. No "Sure, here's...". Just the data.

ANSWER: [one line]
WHY: [max 20 words]

That format alone saved me about 40% on output tokens across my agent.

3. Match the model to the task

Sending every request to Opus is like taking an Uber Black to the bodega.

Here’s my routing logic now:

• Haiku 4.5: classification, simple extraction, “is this email spam,” autocomplete-style stuff. Costs ~$1/M input.

• Sonnet 4.6: 90% of real work. Coding, summarization, agent steps. ~$3/M input.

• Opus 4.7: only when reasoning is genuinely hard. Complex multi-step planning, novel problems, the stuff where a wrong answer is expensive.

I built a tiny router that uses Haiku to classify the incoming request, then picks the right model. The router itself costs almost nothing. My monthly bill dropped 58% the week I shipped it.

If you’re spending more than $500/month on LLM APIs and you’re not routing, that’s your homework.

4. Strip your system prompt to the bone

Most system prompts are 3x longer than they need to be. People paste in role descriptions, ten rules, four examples, edge cases, a tone guide, then wonder why their bill is high.

Try this. Take your current system prompt. Delete every sentence. Add them back one at a time and run the eval. Stop when quality plateaus.

I did this with a writing assistant prompt. Started at 1,800 tokens. Got it down to 340 with no measurable quality drop. Multiplied across 50,000 daily calls, that’s millions of tokens a day.

Words to delete first: “please,” “make sure to,” “it’s important that,” “remember that,” “you should always.”

The model knows. Just tell it the rule.

5. Don’t dump full chat history. Summarize it.

This is the silent killer in agent loops.

Default behavior: every turn, you send the entire conversation history. By turn 20, you’re sending 30,000 tokens of history just to ask “what about the second one?”

Fix: after every N turns (I use 5), summarize the older messages into a compact paragraph and replace them. Claude Code does this automatically with /compact. You can build the same thing into your app.

if len(messages) > 10:
    summary = summarize(messages[:-4])  # cheap Haiku call
    messages = [{"role": "system", "content": f"Earlier context: {summary}"}] + messages[-4:]

One call to a cheap model saves you thousands of tokens on every subsequent expensive call.

6. For RAG: retrieve less, retrieve better

Most RAG apps retrieve the top 10 chunks “just in case.” That’s lazy and expensive.

Three things that worked for me:

• Cut to top 3 chunks. If your retrieval is good, 3 is enough. If it’s not, fix retrieval, don’t paper over it with more chunks.

• Re-rank before passing. Use a cheap re-ranker (Cohere has one, or just use Haiku) to pick the best 3 from the top 10. Then send only those 3 to the expensive model.

• Chunk smaller. 256 to 512 tokens for factoid stuff. 1,024 only when you genuinely need flowing context. Bigger chunks mean more wasted tokens per retrieval.

I cut my RAG input tokens by 64% using just these three. Answer quality went up, not down, because the model wasn’t drowning in irrelevant text.

7. Use structured outputs instead of “respond in JSON”

When you ask “respond in JSON format,” the model often adds preamble. “Sure, here’s the JSON:” then the actual data. Wasted tokens, plus you have to parse it.

Use the API’s structured output feature instead. OpenAI calls it response_format. Anthropic uses tool definitions. The model is constrained to the schema. No preamble. No closing remarks. Just the data.

Bonus: it almost never fails to parse anymore.

8. Few-shot only when zero-shot fails

Few-shot examples are powerful and expensive. A 4-example prompt can easily add 1,500 tokens to every single call.

My rule: try zero-shot first. If quality is bad, try one example. If still bad, two. Don’t lead with five examples because a tutorial said so. Each example is tax you pay forever.

When you do use them, put them inside a cached block (see tip #1). That way you only pay full price once, not per call.

9. Paste the text. Don’t upload the PDF.

This one shocked me when I learned it.

When you upload a PDF to Claude (via the API, Claude Code, or even the chat UI), every single page gets converted into an image and processed as a mix of visual + text tokens. One PDF page costs you 1,500 to 3,000 tokens. Even if the page only has 200 words on it.

That same 200 words pasted as plain text? Around 250 tokens.

So you’re paying a 6x to 12x markup just for the privilege of using a PDF.

Here’s why. PDF is a layout format, not a text format. The model has to OCR each page, recognize fonts and columns, parse tables, and reason over the visual structure. All of that costs tokens. If you don’t need the layout (charts, diagrams, signatures, page-position context), you’re paying for nothing.

Real example. I had a 42-page research report I was iterating on with Claude. Uploaded as PDF: roughly 84,000 tokens per call. I ran pdftotext on it once, copied the output into the terminal, and pasted the clean text into the chat instead. Same back-and-forth: about 9,500 tokens per call. That’s an 88% cut on a single document.

The workflow:

# macOS / Linux: dump PDF to stdout, copy from terminal
pdftotext your_doc.pdf -

# Or just ask Claude Code:
# "Read ~/Downloads/report.pdf and print the cleaned text"

When to actually keep the PDF: if you need the visual stuff. Charts, scanned signatures, diagrams, complex tables where layout matters. For everything else (text reports, papers, transcripts, PRDs), paste the text. Your bill will look very different.

This same trick applies to Word docs, slides, and spreadsheets. Convert to plain text or markdown first. Less container overhead, fewer tokens, cleaner reasoning.

Share

A copy-pasteable audit prompt

If you don’t know where your tokens are going, paste your current system prompt + a sample user message into Claude and run this:

Audit this LLM prompt for token waste. List:

1. Redundant or repeated instructions I can delete
2. Vague phrases like "please" or "make sure to" that don't change behavior
3. Examples that could be removed without quality loss
4. Static parts I should mark for caching
5. Estimated token count before and after your suggested edits

Be ruthless. I'd rather lose words than money.

I run this on any prompt that’s part of a high-volume call. Usually saves 30 to 50% on the spot.

The mindset shift

Most people treat tokens like they’re free. They’re not. Every word in your prompt is a small recurring bill.

Once you start thinking in cost-per-call instead of cost-per-month, everything tightens up. You delete fluff. You cache. You route. You set max_tokens.

And the wild thing? Your apps usually get faster and better when you do this. Less noise in the prompt means cleaner reasoning.

If you are finding value in my AI action letters, please consider subscribing. It keeps me going.

If you know someone who’s burning cash on LLM APIs and doesn’t know why, please share this with them.

That’s it from me today.

Till next time. Stay tuned as I will share best resources from both my Harvard and Google networks to bring the best to you. Let’s up skill together. Aspyre higher!

Yesterday a friend texted me asking if she should apply to the Perplexity AI Research Residency. She doesn’t have a CS degree. She studies cognitive science.

I told her: that’s exactly the point.

A lot of people see “AI research residency” and assume the bar is a PhD from Stanford, three NeurIPS papers, and a GitHub full of transformer implementations. So they don’t apply.

That’s the wrong read.

The Perplexity residency pays $220K a year. It’s three months full-time in San Francisco. You walk away with real research, real connections, and a credential that opens doors at every AI lab in the Bay.

And the program is explicitly built for people who don’t fit the traditional AI mold. Physicists. Quants. Cognitive scientists. Mathematicians. Engineers from adjacent fields.

If that’s you, here’s how to apply without leaving anything on the table.

A quick note for my international readers: Some of you may have applied to the H-1B lottery. If you are stressed about it or this is your last chance, please explore H-1B alternatives, such as EB-1A or O-1 visas, especially if you have a STEM background.

I’ve partnered with Manifest Law to get you a free 15-minute consultation. These are legitimate paths that many people don’t know about. If this applies to you, it’s worth a conversation: Book your free consultation here.

What the residency actually is

Three months. Full-time. In-person, SF or Palo Alto.

Compensation is $220K annualized, which works out to around $55K for the term.

Rolling admissions, so the earlier you apply, the better your shot. But there’s a catch. Only apply when you can actually start within 8 weeks of acceptance. This program moves fast.

The official program page is at perplexity.ai/hub/ai-research-residency and the live job posting is on their Ashby careers site.

The cover letter (most people get this wrong)

The cover letter is where most candidates kill their chances before anyone reads their resume.

Here’s why. Most cover letters open with: “I’m a recent graduate passionate about AI...” and then list classes they took. That’s a template. Templates lose.

Perplexity is explicit on their hiring page. They want to know how your background outside of AI adds value. Frame your non-AI experience as the feature, not the bug.

Some tactical moves:

1. Lead with the transferable lens.

Don’t open with “I love Perplexity.” Open with what you bring. A physicist brings systems thinking. A quant brings probabilistic rigor. A cognitive scientist brings mental models of human reasoning. A linguist brings deep priors on how language fails.

Pick the one that fits you. Make it the first sentence.

2. Write a separate research statement (the cheat code).

Their application has an optional 1-page research statement. Most applicants skip it because it’s optional.

Don’t skip it.

This is the single biggest differentiator in the application. It shows you’ve thought about what you’d actually do during the residency. It separates you from the pile.

A good research statement does three things:

• Names a specific open problem (search reliability, hallucination in citations, multi-step retrieval, agentic browsing)

• Explains why it matters for Perplexity’s mission

• Sketches how you’d investigate it in 3 months

3. Show you’ve used the product.

Use Perplexity daily for at least a week before you apply. Read research.perplexity.ai. Their actual published work tells you what they care about. Reference one of their papers in your cover letter or research statement.

Almost no one does this.

4. Be specific about timing.

If you can start within 8 weeks, say so explicitly in the cover letter. They’re not hiring abstract candidates. They’re hiring someone to walk in the door in two months.

5. Skip the buzzwords.

No “passionate about pushing the boundaries of AI.” No “cutting-edge research.” Just say what you’ve done and what you want to do next. Perplexity researchers are pattern-matching on the same generic AI fluff you are. Don’t be that pile.

Cover letter template (use as a starting point, not a copy-paste)

Dear Perplexity Research Team,

I'm a [your specific background, e.g. computational neuroscience PhD candidate /
quant researcher at X / theoretical physicist working on Y]. My work centers on
[the one technical problem you're known for].

I'm applying to the Research Residency because I want to spend 3 months working on
[specific Perplexity-relevant problem]. In my field, we approach [related problem]
by [technique]. I think the same lens applies to [Perplexity research area you've
identified from research.perplexity.ai].

Three things I'd bring:
1. [A skill from your field, with a concrete example]
2. [A technical capability, with a result]
3. [A research instinct, with a project that shows it]

I've attached a 1-page research statement with a specific question I'd like to
investigate during the residency. I can start within [X] weeks of an offer.

[Your name]

Don’t copy those exact words. Use the structure.

The interview process

Based on candidate write-ups across Substack, Medium, and Glassdoor, here’s the 5-stage process most residency candidates report:

Stage 1: Recruiter screen (~45 min). Background, tech stack, hardest project you’ve shipped, comp expectations.

Stage 2: Technical assessment. Python coding. Array manipulation, string parsing, sometimes a small simulation problem. ML fundamentals. First-principles reasoning.

Stage 3: Research discussion. Walk through a research problem you’ve owned end-to-end. They want to see how you frame, hypothesize, run experiments, and learn.

Stage 4: Collaborative problem-solving. Open-ended, half-baked problem with a potential mentor. No right answer. They want to see how you ask clarifying questions and think under uncertainty.

Stage 5: Cultural fit. Do you align with the mission. Do you take ownership. Do you move fast.

The whole loop wraps in 3 to 21 days.

Practical interview tips

Python is the language. Don’t show up with C++ or Rust. One Medium write-up from a successful candidate said it directly: “Python is king for this interview. Stick to it.”

Talk while you code. Communication is weighted as heavily as the answer. Narrate your thought process. Surface assumptions out loud. State what you’d test if you had more time.

Have a research story ready. Pick one project you’ve owned. Be ready to walk through it for 15 minutes covering: the question, what you tried, what failed, what worked, what you’d do differently. This comes up in stage 3 every single time.

Study these specific topics:

• Transformers: attention, why it works, what breaks at scale

• Embeddings: what they encode, how to evaluate them

• RAG systems: retrieval failure modes, chunking strategies, eval

• Agentic systems: this is Perplexity’s bet. Build something that “does” not “says”

Read 2 or 3 papers from research.perplexity.ai. Be ready to discuss one in detail.

For the collaborative round, ask first. When they hand you an ambiguous problem, don’t start solving. Ask three clarifying questions. Who’s the user. What failure mode are we trying to fix. What constraints am I working with. This is the round where most candidates lose because they jump in.

Have a Perplexity critique ready. They love this. Use the product. Find one thing that’s broken or could be sharper. Say it directly. Don’t be sycophantic. Don’t be brutal. Be honest and specific.

Questions you’ll be asked (prep these word for word)

1. Walk me through a research problem you’ve worked on end-to-end.

2. What’s an open research question in your field that AI could help solve?

3. How would you design an experiment to test [X]?

4. What are the failure modes of your proposed approach?

5. Why Perplexity, specifically? Not OpenAI, not Anthropic, not a lab. Perplexity.

That last one matters. “I love AI” is not an answer. The right answer ties their specific mission (making search reliable, building agentic systems for information work) to what you actually want to build. If you can’t tell the difference between Perplexity and a generic AI lab, you’re not ready to interview yet.

Who should apply

If you have a strong technical background in any field, the answer is probably yes. Even if you’ve never trained a model. Especially if you’ve never trained a model and have a unique lens on a hard problem.

If you’re a CS undergrad who’s done two ML classes and has nothing else on the resume, this one is harder. The whole point of the program is to bring in people the standard AI hiring funnel misses. Lean into what makes you weird.

If you are finding value in my AI action letters, please consider subscribing. It keeps me going.

If you know someone who’s been on the fence about applying to AI roles because they don’t have an AI background, please share this with them. The Perplexity residency is one of the rare programs that explicitly wants them. Apply here.

That’s it from me today.

Till next time. Stay tuned as I will share best resources from both my Harvard and Google networks to bring the best to you. Let’s up skill together. Aspyre higher!

Saturday morning. Chai in hand. A friend forwarded me a Reddit post making the rounds: “I tested 50 secret Claude prompt codes. Most are fake. Here are the 7 that actually work.”

I rolled my eyes. Every other week somebody claims they cracked the model.

I’m going to walk you through each one with a career-upskilling use case. Steal these for your job search, your cold outreach, your interview prep, and your offer negotiation.

Disclosure: There are NO official secret codes, but because people have used them so much, Claude has actually started responding to them

1. /ghost: writes like a human, not a model

Default Claude writes like Claude. Soft openings. Perfect tricolons. “Here’s a draft for you to consider.”

Type /ghost before any writing prompt and it strips all of that. No meta-commentary. No signposting. No “I hope this helps” closer.

Best use case: cold LinkedIn DMs and outreach emails.

I tested it on a LinkedIn message to a Google PM I wanted to chat with:

/ghost Write a 3-line LinkedIn DM to a senior PM at Google asking for a 15-min chat. I'm a marketing analyst trying to break into product. Don't sound like AI.

Without /ghost you get four paragraphs and a “looking forward to connecting.” With /ghost, three sentences, direct ask, sounded like a person.

The PM responded the same evening.

I am not saying the prefix did it. But the message read like a human and that’s the only bar that matters for cold outreach in 2026.

2. /godmode: strip the hedging, get the real take

Claude defaults to balance. Ask whether you should leave your job to do a startup and you get a 600-word “it depends” essay covering both sides.

/godmode removes the both-sides default. You get the answer Claude actually thinks is right.

Best use case: honest career feedback when you don’t want a tour of options.

I tried this:

/godmode I'm a 4-year senior data analyst at a Series C startup. Two offers in hand. Senior Data Scientist at Stripe ($200K base) and Staff Analyst at Coinbase ($240K base). Which should I pick?

Without /godmode it gives a comparison framework. With /godmode it picks Stripe, explains exactly why (title trajectory, learning ceiling, alumni network, the $40K delta is recoverable in 18 months at Stripe but the brand is not), and ends with the specific risks of each path.

That’s the answer you actually came for.

Use it when you want a take, not a tour.

3. L99: at the END of a prompt, for expert depth

This is the one most people get wrong. L99 goes at the END of the prompt, not the beginning. It signals “give me the expert-level answer, not the summary.”

I tested it on a real PM interview prep question:

Walk me through how you would design a feature to reduce churn for a paid news subscription product. L99

Without L99 you get a clean 5-step framework you could find on any blog.

With L99 you get: the framework PLUS the specific churn metric to anchor on, the three segmentation cuts most candidates miss, the trade-off between win-back campaigns vs prevention features, and the answer that would actually score well at the Google PM interview bar.

The depth shift is not subtle.

Use it for: interview prep, technical questions, decisions where the wrong answer is expensive.

4. ARTIFACTS: make Claude BUILD the thing, not describe it

This one is the most underused code on the list. ARTIFACTS is not just “give me deliverables in text.” It triggers Claude’s artifacts feature, which builds a live, working interactive tool inside the chat. Working code. Working dashboard. Working calculator.

For career upskilling, this is the unfair advantage nobody is talking about.

I asked Claude:

ARTIFACTS Build me a job application tracker dashboard. I want to log company, role, JD URL, application date, status, contact info, follow-up dates. Show me a Kanban view by stage. Let me filter by company.

Got back a fully working React dashboard. Right inside the chat. I could add applications, drag cards between stages, filter by company. Eight minutes from prompt to functional tool.

I have used it this month to build:

• Salary negotiation calculator: compares total comp across offers (base, bonus, equity vesting, sign-on, refresh)

• Cold-email response tracker: logs reply rates by template, tells me which subject line is winning

• 90-day learning roadmap visualizer: weekly check-ins, tracks completion against my target role

• Mock interview question bank: randomizes behavioral questions by category

If you have ever wanted a custom job-hunt tool but didn’t want to learn React, this is your moment.

5. OODA: for career decisions on a short clock

OODA is a real military framework. Observe, Orient, Decide, Act. Claude knows it cold.

Prefix any high-stakes decision with OODA and you get four structured sections: what is actually happening, how to read it, what to choose, what to do this week.

A friend of mine was in this spot last month:

OODA Two offers, decide by Friday. Offer A: PM role at a YC-backed Series A, $180K base, no equity benchmark, founder is ex-Stripe. Offer B: Senior PM role at Microsoft, $220K base, full benefits, well-defined ladder, no PM mentor.

What it returned:

• Observe: laid out the real signals (founder pedigree, ladder definition, comp delta, optionality cost)

• Orient: framed it as “career capital vs career stability” and pushed her on which she actually needs in the next 3 years

• Decide: picked the YC startup, because she had asked specifically for fast learning, not stability

• Act: gave her three exact questions to ask the founder before signing, plus the salary counter to negotiate

She signed Tuesday.

Use OODA when you’re stuck and the clock is short. Postmortem on a bad interview. Deciding whether to switch teams. Evaluating offers. Anywhere the structure of the decision is more useful than the decision itself.

The honest take nobody is saying out loud

Here’s what I noticed after 3 hours of testing.

These codes are not secret. They are not revealing hidden capabilities. The model is not hiding anything from you.

They work because they are compressed shortcuts for “be specific about what you want.” L99 is a 3-character version of “give me an expert-level answer with depth, not a summary.” OODA is a 4-letter version of “structure this as Observe, Orient, Decide, Act.”

You can write the full instructions in plain English. You will get the same answer.

So why use the codes?

Because you’ll actually do it.

Typing /ghost is 6 keystrokes. Typing “remove all meta-commentary, soft openings, signposting, and AI-tells from your response” is 90 keystrokes. The codes are not magic. They are friction reduction. They make you prompt better consistently because they are easy.

That’s the real lesson of the Reddit post. The “secret prompts” debate misses the point entirely. The codes are not the moat. Specificity is the moat. The codes just make specificity cheap.

Steal the 5 above. Use them this week. Write the longer English versions when you forget the codes.

Either way, your prompts will get sharper. And in 2026, prompting is the real career skill.

If you are finding value in my AI action letters, please consider subscribing. It keeps me going.

If you know someone hunting for a job or stuck on a career decision this week, please share this with them. The OODA code alone is worth forwarding.

That’s it from me today.

Till next time. Stay tuned as I will share best resources from both my Harvard and Google networks to bring the best to you. Let’s up skill together. Aspyre higher!

Two weeks ago a PM friend of mine pinged me on Whatsapp at 11pm.

She said, “Abhijay, I’ve been paying for Claude Pro for nine months, and I see no difference compared to ChatGPT. Am I missing something?”

I laughed. Because I did the exact same thing for almost a year.

Claude today isn’t one tool. It’s a suite of five, each built for a totally different job.

Most people are using one of them. The other four are just sitting there.

So today I want to walk you through all five, with 3 copy-pasteable prompts for each. No fluff. Stuff you can run today.

Let’s go.

The 5 Claude products at a glance

• Claude Chat: your thinking buddy. Use it when you’re stuck.

• Claude Design: your in-house designer. Give it a brand link, get landing pages.

• Claude Code: your dev team on tap. Build the tool you wish existed.

• Claude Projects: your second brain. Upload context once, never repeat yourself.

• Claude Cowork: your virtual assistant. Cleans folders, writes reports, does the grunt work.

Five different tools. One subscription. Most creators are using 20% of what they’re already paying for.

Now the prompts.

1. Claude Chat (your thinking buddy)

This is the regular Claude.ai chat window. The one most people already know.

But here is the trap: most folks use it like Google. Ask a question, get an answer, close the tab.

The real value shows up when you treat it as a thinking partner. Stuck on a decision. Stuck on a strategy. Stuck on how to phrase a hard email. That’s when Chat earns its keep.

3 prompts that punch above their weight:

Decision pressure test

I'm trying to decide between [Option A] and [Option B].
Context: [paste the situation, what's at stake, what I've already tried].

Don't tell me which one to pick. Instead:
1. Ask me 5 sharp questions that surface what I actually care about.
2. After my answers, lay out the real tradeoff in one paragraph.
3. Tell me the option I'm probably leaning toward and why I might be wrong about it.

The "explain it back to me" loop

I just read [topic / article / paper / book chapter].
Here is my summary in my own words: [paste your understanding].

Tell me:
1. What I got wrong.
2. What I oversimplified.
3. What I missed entirely.
4. The 1 idea from this that I should actually use this week.

Be ruthless. Pretend you're reviewing a PM's first draft.

The 60-minute strategy sprint

I have 60 minutes to make progress on [specific goal].
Constraints: [time, budget, who's available, what I cannot do].
Today's energy: [low / medium / high].

Build me a 60-minute plan with exactly 3 high-impact tasks in order.
For each task, tell me:
- The exact output I should have at the end.
- What I'll skip if I run out of time.
- The single decision I need to make first.

2. Claude Design (your in-house designer)

This is the one that surprised me the most.

You drop a website link or a brand brief, and Claude builds you actual landing pages. Following your existing brand. With copy that doesn’t sound like every other AI-generated landing page. Yes, you can give it your existing Figma files too.

I tested this on my own newsletter. I gave it my Substack URL and asked for 3 landing page variants for a paid course idea. It gave me 3 fully designed HTML pages in about 4 minutes. They were better than the $400 freelancer I hired last quarter. Not exaggerating.

3 prompts to actually get something usable:

3-variant landing page test

My website: [URL]
Product / offer: [what you're selling]
Price: [number]
Target reader: [who they are, what keeps them up at night]

Build me 3 landing page variants:
1. Pain-led. Open with the problem.
2. Result-led. Open with the outcome.
3. Story-led. Open with a personal or customer anecdote.

Each page under 600 words. Match the colors, typography, and tone of my existing site.
At the top of each, leave a comment block telling me which audience this variant is best for.

The single-screenshot pitch slide

Design a single landing page that doubles as a screenshot for X and LinkedIn.
Product: [your thing]
Audience: [who]

Required elements:
- Hero line in 8 words or less
- Sub-hero in one sentence
- 3 proof points (with real numbers, not vague claims)
- 1 social proof quote
- 1 CTA button

High contrast. Big typography. The whole page must read in under 7 seconds.

Competitor reverse-engineer

3 sites I admire: [URL 1], [URL 2], [URL 3]

Audit each one in 3 bullets: structure, strongest element, what's underdone.

Then build me a landing page for [my product] that takes the strongest element from each.
At the top of the file, comment which element you took from where, and why.
Don't copy. Translate.

Subscribe now

3. Claude Code (your dev team on tap)

Claude Code is the CLI tool. Lives in your terminal. Reads your files, writes new ones, runs commands, ships working software.

This is the one I use most as a creator. Not because I’m a developer. Because I have a long list of “I wish a tool existed that...” problems, and Claude Code can build most of them in under an hour.

Last month I asked it to build me a script that scrapes my top 10 Instagram reels every week and ranks them by an outlier score. It works. I run it every Monday. A freelancer quoted me $2,400 for the same thing.

3 prompts that actually work for non-developers:

The "tool I wish existed" prompt

I want a [tool / script] that does [specific job].
Inputs: [what I'll feed it, where the data lives].
Outputs: [exactly what I want back, in what format].

Constraints:
- I am not a developer. Pick the simplest stack possible.
- Explain each step in plain English as you go.
- No fancy frameworks. Boring is good.
- When done, give me the exact 1-line command to run it.

The folder audit

Audit this folder: [path].

Tell me:
1. Which files haven't been touched in 60+ days.
2. Which scripts look like duplicates or near-duplicates.
3. Which files are doing real work right now.
4. What I can safely archive.

Don't delete anything. Just give me a 5-line summary I can read in 30 seconds, then a longer breakdown I can scan if I want.

The "automate my Monday morning" prompt

Every Monday I do this manually:
1. [step 1]
2. [step 2]
3. [step 3]

It takes me about [X minutes / hours].

Build me a CLI command I can run as `monday-prep` that automates the boring 80%.
Add a confirmation step before anything irreversible.
At the end, print a 3-bullet summary of what it did.

4. Claude Projects (your second brain)

Projects is the feature most creators sleep on.

Here’s what it does. You create a Project. You upload your past chats, transcripts, brand docs, customer feedback, whatever lives in your head. From then on, every conversation in that Project remembers all of it. Forever.

I have a Project for my Substack. It has my last 80 articles, my voice guide, and a folder of reader DMs uploaded. When I draft a new piece, I just open that Project and it already knows my style. No more pasting “write in my voice” every time.

3 ways to set up a Project that actually pays off:

The voice memory dump

I'm uploading 10 of my best [posts / emails / scripts].

Read all of them. Then build a "voice profile" doc with:
1. My top 5 recurring phrases.
2. My average sentence length and how often I break the rhythm.
3. Words and constructions I never use.
4. My opener pattern.
5. My closer pattern.

Save the result as the project's reference file. Every future draft in this project should be checked against it.

The customer brain

I'm uploading every piece of feedback, DM, and review I've collected in the last 6 months.

Read all of it. Tell me:
1. The top 3 things people actually buy from me. Not what I think I sell.
2. The top 3 objections that keep coming up.
3. The exact language they use to describe the result they got.
4. The 1 customer quote I should put on my homepage tomorrow.

The weekly debrief

Every Friday, I'll paste my week into this project:
- 3 wins
- 3 losses
- Open questions

Job: track patterns over time. Don't summarize each week. Look for what repeats.
Every 4 weeks, generate a "what I keep doing wrong" report with specific examples from the previous month.

Don't be polite. Tell me what my friends won't.

5. Claude CoWork (your virtual assistant)

This is the newest one in the suite, and the one I’m still figuring out.

Claude Cowork takes over the grunt work on your machine. Cleaning up folders. Auditing your desktop. Pulling numbers from spreadsheets. Renaming 200 files in a consistent format. The stuff a VA would do, except it happens in 90 seconds.

I asked it to organize my Downloads folder last Tuesday. 14 months of chaos. It sorted everything into 6 folders by type, flagged 22 duplicates, and gave me a markdown summary of what it found. That alone saved me an afternoon.

3 prompts that earn their keep:

The desktop sweep

Look at my desktop and Downloads folder.

1. Group all files by type: PDFs, images, screenshots, code, video, random.
2. Flag anything I haven't opened in 60+ days.
3. Move stale files into a folder called "archive-[today's date]".
4. Find duplicates and near-duplicates. Keep the newest one.
5. Hand me a 5-line summary of what you did and what's left.

The Monday metrics report

Folder: [path with your CSVs, exports, dashboards].

Pull these numbers for last week vs the week before:
- [metric 1]
- [metric 2]
- [metric 3]

Flag anything that moved more than 15% in either direction.
Output one Markdown report I can paste straight into my team Slack on Monday at 9am. Headlines on top. Detail below.

The PDF triage

Read every PDF in [folder].

Sort them into 3 buckets:
1. Needs action this week.
2. Reference only. File and forget.
3. Safe to delete.

For bucket 1, give me a one-line summary of what action it needs and the deadline (if any).
For bucket 2, tag each with 1-3 keywords so I can find them later.
For bucket 3, just list filenames so I can scan and approve.

The thing nobody tells you

You’re not paying for one product. You’re paying for a suite.

Most creators use 20% of what’s in front of them. Then they go shopping for another tool to solve a problem that the tool they already have can do.

Here’s the move. Pick the one tool from this list you’ve never opened. Open it tomorrow. Run one prompt from this article. See what happens.

You’ll know in 10 minutes if it changes your week.

If you are finding value in my AI action letters, please consider subscribing. It keeps me going.

If you know someone who pays for Claude and only uses the chat window, please share this with them. They are leaving most of the value on the table.

Check out this repository for 50+ skills.

That’s it from me today.

Till next time. Stay tuned as I will share best resources from both my Harvard and Google networks to bring the best to you. Let’s up skill together. Aspyre higher!

Last Tuesday, I sat down to write a LinkedIn post about the AI tools I use every day. I typed seven words into Claude. “Write me a LinkedIn post about AI.”

You can guess what came back.

A grey, average, every-other-creator-on-the-internet kind of post. The “AI is changing the game” type. Generic. Forgettable.

I’ve been there a hundred times. Maybe you have too.

I used to blame the model. Surely Opus 4.6 should be smarter than this. But the model wasn’t the problem. My prompt was.

So I learned to prompt in 3 levels. Same model. Same task. 10x better output, every single time.

Btw, if you find value in my articles, consider pledging a amount. It feels good to know someone is willing to pay for what I share :)

Subscribe now

Here’s the stack I now use for everything. LinkedIn posts, Substack drafts, exec memos, even my resume edits.

Move 1: Just ask

This is where most people start. And stop.

You type one sentence. Claude guesses what you want. It pulls from priors and hands you the average of every blog post on the internet.

No audience. No angle. No voice. Nothing for the model to anchor on.

The output is fine. But fine is the enemy of good.

Move 2: Make Claude interview you

Two words flipped this whole thing for me.

“Don’t draft yet.”

Now your prompt looks like this:

I want to write a LinkedIn post about the AI tools I use daily.
Don't draft yet. Interview me first.
Ask me about angle, audience, and what I actually want to say.

Watch what happens. Claude turns into your editor instead of your ghostwriter. It pauses. It asks questions. It pulls things out of your head that you forgot you knew.

Now it writes your post. Not the average.

This single tweak takes you from 30% effort to 70% quality. No exaggeration.

Move 3: Hand it your brain

Move 2 is good. Move 3 is unfair.

Before you ask Claude anything, drop in your context files. The stuff that makes you, you. Three files do most of the work for me:

• about-me.md: my background, what I do, what I care about

• how-i-write.md: my voice, my favorite phrases, my rhythm

• copy-rules.md: the words I never use (looking at you, delve)

Then your prompt is the full stack:

I want to [your task].

First, read every attached file in full before you respond.

Don't draft yet. Instead, interview me. 
Ask the questions you'd ask if I'd hired you.

Claude now starts the chat already knowing your story, your voice, your rules. Then it gets to ask. By the time it drafts, it’s basically writing as you.

Why this stack works

Every prompt that actually lands does 3 things. This stack does all three.

1. Hands over raw material. Files, examples, half-baked notes. Without them, Claude works from priors. With them, it works from you.

2. Buys time before writing. “Don’t draft yet” forces the model to think. It surfaces hidden assumptions before they become 800 wasted words you have to cut.

3. Makes it interview you. Questions catch the misread up front. You always know more than the prompt captures. Let it pull that out of you.

That’s it. Three moves. Compounds fast.

My setup so I never re-do this

Here’s what I actually do day to day. I save the template above as a Claude Project. I attach the 2-3 context files once. Done.

Now every chat I open already starts at Move 3. I just type the task and go.

Same model. Sharper prompt. 10x the output.

If you want a head start, the three files I attach to every chat (about-me, anti-AI-style, copywriting) are the ones I’m sharing on LinkedIn this week. Comment PROMPT on my carousel and I’ll send them to you.

If you are finding value in my AI action letters, please consider subscribing. It keeps me going.

If you know someone who is frustrated that their AI outputs sound generic, please share this with them.

That’s it from me today.

Till next time. Stay tuned, as I will share the best resources from both my Harvard and Google networks to bring you the best. Let’s upskill together. Aspyre higher!

Manually.

Copy-paste resume. Tweak cover letter. Click LinkedIn Easy Apply. Refresh inbox. Repeat.

You know what I got back? Automated rejections. No replies. Not a single interview request.

I remember lying awake thinking: I just took a massive loan to come to the US. I have a Harvard MBA. And I can’t even get someone to read my resume.

That was before the AI boom.

Today, I’m a Product Manager at Google — and I haven’t manually submitted a job application in years. Because I built a system that does it for me.

Every morning at 7am, while I’m asleep, an AI agent wakes up, scrapes the freshest job postings on LinkedIn, rewrites my resume for each one, generates a polished PDF, and emails me a complete set of tailored applications. By the time I have my coffee, the work is done.

This article is the full breakdown. Every tip, every prompt, and the exact N8N workflow I use, including the free JSON template you can copy and run today.

Here’s everything you need to get started:

• Free N8N workflow JSON template here

• Self-host N8N for $7/month here (don’t pay $24/month on n8n’s website!)

• Set up Supabase here

Each article takes a lot of research, workflow engineering, and drafting. I request you to pledge an amount to support my work so I can continue the series.

Why everything you’re doing right now isn’t working

If you’re using LinkedIn Easy Apply, I need you to stop.

I know it feels productive. It’s not.

Here’s what’s actually happening: by the time you see a job posted on LinkedIn and hit Easy Apply, there are already 100+ applicants ahead of you. The ATS system filters out 70%+ of them before a human ever reads a single resume. The ones that survive are the ones with exact keyword matches to the job description. Your generic resume — the one you sent to 50 other companies this week — doesn’t have those keywords.

This is the whole game. Not applying more. Applying smarter, faster, and to jobs that actually have room for you.

Here’s the system that does that.

The 10 AI strategies that actually work

I’m splitting these into three phases: finding the right jobs, applying, and the post-offer stage.

Phase 1: Finding Jobs

Tip 1: Only apply to jobs posted in the last 24 hours

This is the single highest-leverage change you can make right now.

Most people are competing for jobs that are 5-7 days old. Those have hundreds of applicants. Jobs posted in the last 24 hours have almost none.

Here’s how to find them consistently. Open Claude, enable the Apify connector, and use this exact prompt:

“Find me the most recent [your target role] jobs in [city] posted in the last 24 hours. Use Apify to scrape LinkedIn jobs and return them as a spreadsheet with company, role, link, date posted, and an ATS match score against my resume: [paste resume].”

Claude will generate a live spreadsheet filtered to your profile, with an ATS match score so you know your chances before you even apply.

Later in this article, I’ll show you how to automate this so it runs every morning without you touching it.

Tip 2: Find jobs directly from hiring managers

Most people apply through job boards. The people actually getting hired are going directly to the source.

Hiring managers regularly post on LinkedIn saying “we’re hiring.” Almost nobody finds these posts because they’re not listed on any job board.

Use this prompt in ChatGPT, Claude, or Gemini:

“Search LinkedIn for posts from hiring managers in [your target role] at [target companies] posted in the last 7 days. Return: name, company, LinkedIn URL, and a summary of what they’re hiring for.”

DM them directly. You will almost always get a reply. This was the move that changed my entire job search.

Phase 2: Applying

Tip 3: Stop sending the same resume to every job

Every job description contains the exact keywords the ATS is scanning for. Your resume needs to match them.

This doesn’t mean lying or inflating your experience. It means surfacing the right words.

Every time you find a role you want to apply to, do this:

“Here is my resume: [paste]. Here is the job description: [paste]. What keywords are missing from my resume that appear in the job description? What should I change to pass an ATS filter?”

Takes 3 minutes. Does more for your hit rate than anything else.

Tip 4: The N8N automation that does all of this while you sleep

This is the main event.

I built an N8N workflow that fully automates the job search and application process. Here’s exactly what it does, every morning at 7am:

• Pulls your resume from Google Drive

• Scrapes the 10 most recent matching jobs from LinkedIn (filtered to last 24-48 hours)

• Checks Supabase to filter out jobs already processed — no duplicates, no wasted credits

• For each new job: rewrites your resume using an AI agent (Google Gemini) to match the exact keywords and requirements of that specific role

• Compiles the optimized resume into a clean LaTeX PDF

• Uploads each PDF to Google Drive and makes it shareable

• Emails you a summary with links to every tailored resume, organized by company and role

You wake up to something like this in your inbox:

Click through. Each link goes to a polished PDF, already tailored to its role. Apply. That’s the whole morning routine.

Get the free N8N workflow JSON template here.

How to set this up (even if you’ve never used N8N)

N8N is the automation tool that runs the whole system. You self-host it so you control everything and pay a fraction of the cloud price.

Step 1: Set up N8N on Hostinger ($7/month)

Go to hostinger.com/abhijay and pick the KVM 1 VPS plan (the cheapest option — it’s enough for this workflow).

Once your VPS is live, go to the AI Assistant in your Hostinger dashboard and type: “Install N8N with Docker.” It walks you through the whole thing. No command line expertise required.

When it’s done, your N8N instance will be running at your VPS IP address. Open it in a browser and you’re in.

I’m using Hostinger specifically because of the AI VPS Assistant — it’s genuinely useful when you’re configuring things for the first time and get stuck. Ask it anything.

Step 2: Import the workflow

In N8N: File → Import from JSON → paste the template from the Google Doc above.

Step 3: Fill in the Workflow Configuration node

There’s one node called “Workflow Configuration” — it’s the only node you need to edit. Fill in these 5 fields:

• resumeFileId — the ID from your Google Doc resume URL (the long string between /d/ and /edit)

• jobSearchQuery — e.g. “AI Product Manager San Francisco”

• userEmail — where you want to receive the morning summary

• apifyToken — get a free token at apify.com

• supabaseUrl + supabaseKey — from your Supabase project dashboard (see Step 4)

Step 4: Set up Supabase for deduplication (free)

Without this, the workflow will keep re-processing the same jobs every morning and burning your API credits.

Sign up at supabase.link/N3ar6hP — the free tier is more than enough.

Create a new project. Go to Table Editor → New Table. Name it jobs. Add these columns:

• job_url (text, primary key)

• job_title (text)

• company (text)

• processed_at (timestamptz)

Go to Project Settings → API and copy your Project URL and anon key. Paste them into the Workflow Configuration node.

That’s it. The workflow will now track every job it processes and skip anything it’s already seen.

Step 5: Connect your credentials

In N8N, you’ll need to connect:

• Google Drive (OAuth — connect your Google account)

• Gmail (OAuth — same Google account, or a different one)

• Google Gemini API (free — get a key at aistudio.google.com)

Each node that shows a credential error will prompt you to connect — it’s a guided flow.

Step 6: Activate

Hit the toggle in the top right of N8N to activate the workflow. It’ll run automatically every day at 7am.

First run, I’d recommend clicking “Execute Workflow” manually to verify everything’s working before you set it loose overnight.

Tip 5: Use AI to write your warm referral email

The highest conversion rate in job searching isn’t applying online. It’s a warm referral.

If you know anyone at a company — a college connection, a former colleague, someone you met at a conference — they can send one message to the hiring manager and jump you to the top of the pile.

Most people don’t ask because they don’t know how without it being awkward. Here’s the prompt:

“I’m applying for [role] at [company]. I know [person’s name] who works there. Write me a 3-sentence message asking for a warm referral — keep it friendly, specific to my background, not transactional.”

Send it. The worst that happens is they say no. The best is you skip the ATS entirely.

Tip 6: Find hiring manager email addresses automatically

There’s another N8N automation I run for outbound outreach. You give it a search query like “Product Managers working at Stripe on payments infrastructure” and it:

• Runs a Boolean search to find their LinkedIn profiles

• Uses Hunter.io or Apollo.io to verify their email addresses

• Writes a personalized cold email based on your background and their work history

• Drops the draft directly into your Gmail drafts folder, ready to review and send

You don’t send anything automatically. You review each draft and hit send when you’re happy with it. But the research and writing is done

Access it here.

Tip 7: Apply to adjacent roles

The market is hard right now. Big tech is laying off. Startups are laying off. AI is the stated reason for most of it.

If you’re only applying to your exact target title, you’re narrowing your window unnecessarily. Once you’re inside a company, internal transfers are straightforward — especially in big tech.

• Applying for Product Manager? Also apply for Product Marketing Manager, Program Manager, Strategy & Operations.

• Applying for Data Analyst? Also apply for Data Activation Analyst, Growth Analyst, Business Intelligence Analyst.

• Applying for Software Engineer? Also apply for Solutions Engineer, Technical Program Manager, Developer Advocate.

The N8N workflow makes this easy — just run multiple job search queries in parallel.

Subscribe now

Phase 3: Post-Offer

Tip 8: Use AI to prep for every interview

Before any interview, I generate a custom prep guide. Here’s the exact prompt:

“Generate an HTML interview prep page for [role] at [company]. Include 10-15 questions — behavioral and technical — based on this job description: [paste]. Also research the hiring manager’s background using this LinkedIn profile URL: [paste]. Tailor questions to their work history and priorities.”

Claude will give you a full page with questions, suggested answer frameworks, and context on your specific interviewer. I’ve gotten interviewers to ask me questions almost word-for-word from the prep guides I generate this way.

Tip 9: Negotiate your salary with data, not guesses

The worst thing you can do in a salary negotiation is name a number without market data backing it up.

Companies in Silicon Valley offer large signing bonuses specifically because it’s a one-time cost — they’d rather give you $50K upfront than raise your base by $10K/year compounding. If you don’t know this, you’ll optimize for the wrong thing.

Use this prompt before any negotiation:

“I’m negotiating a [role] offer at [company] in [city]. What’s the current market rate for base salary, signing bonus, and equity? Give me P25, P50, and P75 benchmarks. What’s a realistic counter-offer script?”

Walk in with data. Don’t leave money on the table.

Watch the entire video on YouTube:

The shift that changed everything

When I was manually applying to 100+ jobs a day, I felt productive. I was busy.

But busy isn’t effective.

The shift happened when I stopped thinking about job searching as something I do and started treating it as a system I build once and let run. That’s the real unlock that AI gives you — not just doing the same thing faster, but building infrastructure that works whether you’re awake or not.

If you’re in a job search right now, start with one thing: set up the N8N workflow. Even if you only use it to surface fresh job postings every morning, it changes how you approach the entire process.

Here’s everything you need to get started:

• Free N8N workflow JSON template here

• Self-host N8N for $7/month here

• Set up Supabase for free

And if you want 5 more AI job search tips that didn’t make it into this article, comment “guide” and I’ll send them over personally.

Share this with a friend who is still manually applying to jobs on LinkedIn.

Aspyre higher!

— Abhijay

Abhijay is a Harvard MBA, Product Manager at Google, and creator of the AI Action Letter — read by 180,000+ people. Follow on LinkedIn and Instagram.

Let me be honest with you.

I used to roll my eyes at AI certs. Most of them are just vocab quizzes. “What is an LLM?” “Which of these is a transformer?” The kind of thing you could pass after two YouTube videos at 2x speed.

Then I sat down with Anthropic’s new Claude Certified Architect: Foundations (CCA-F) exam guide last week. 60 questions. 120 minutes. Pass score of 720 out of 1000. Every single question is a production-grade scenario.

One of the sample questions literally shows you agent error logs. It says: “in 12% of cases, your agent skips get_customer and calls lookup_order using only the customer’s stated name, occasionally leading to misidentified accounts and incorrect refunds.” Then it asks what you’d change.

That’s not a vocab test. That’s a “can you actually ship with Claude” test.

And honestly? It’s about time.

Consider pledging an amount to my newsletter. It helps me offset the cost of putting together the best resources for all of you.

Why this cert matters in 2026

Here’s the thing.

Every company I talk to right now is hiring for “AI engineers.” But nobody can tell you what that means. Is it someone who calls an API? Someone who fine-tunes models? Someone who designs agent systems?

The job title is meaningless because there’s no industry bar for what a competent AI builder looks like.

Until now.

The CCA-F is the first cert I’ve seen that actually defines that bar. And because it comes from the people who built Claude, it carries more weight than any third-party program.

A few things I pulled from the official guide:

• It’s not a knowledge test. It validates whether you can “make informed decisions about tradeoffs when implementing real-world solutions.”

• The target candidate is a solution architect with 6+ months hands-on experience building with Claude APIs, Agent SDK, Claude Code, and MCP.

• The exam guide warns you: “distractors are designed for candidates with incomplete knowledge.” Wrong answers sound reasonable on purpose.

Translation: you can’t fake this one.

The 5 domains (and why the weights matter)

The exam covers five domains. The weights tell you exactly where to spend your time.

• Domain 1: Agentic Architecture and Orchestration (27%). The biggest one. Agentic loops, coordinator and subagent patterns, hooks, task decomposition, session management.

• Domain 2: Tool Design and MCP Integration (18%). Tool descriptions, error handling, tool_choice configurations, MCP server setup.

• Domain 3: Claude Code Configuration and Workflows (20%). CLAUDE.md files, custom slash commands, skills, plan mode, CI/CD.

• Domain 4: Prompt Engineering and Structured Output (20%). Few-shot prompting, JSON schemas, validation loops, batch processing.

• Domain 5: Context Management and Reliability (15%). Preventing context degradation, escalation patterns, error propagation, information provenance.

If you’re short on time, three domains (Agentic Architecture, Claude Code, Prompt Engineering) make up 67% of the exam. Start there. Skip nothing, but weight your hours accordingly.

The 6 exam scenarios (know all six)

Every question on the exam is grounded in one of six scenarios. Four of them show up on your test at random. Which means you have to know all six because you don’t get to pick.

1. Customer Support Resolution Agent. Tools like get_customer, lookup_order, process_refund. Target is 80% first-contact resolution.

2. Code Generation with Claude Code. Refactoring, debugging, documenting with CLAUDE.md, custom slash commands, plan mode.

3. Multi-Agent Research System. A coordinator delegates to specialized subagents: web search, doc analysis, synthesis, report generation.

4. Developer Productivity with Claude. An agent explores codebases, reads legacy systems, generates boilerplate using Read, Write, Bash, Grep, Glob, plus MCP.

5. Claude Code for CI. Automated code reviews, test generation, PR feedback running inside your pipeline.

6. Structured Data Extraction. Pulling information from unstructured docs using JSON schemas, validating output, handling edge cases.

Pro tip. Build at least one toy version of each. I don’t care if it’s 100 lines. The muscle memory of having shipped each pattern is what the exam is actually testing.

One question that shows you how the exam thinks

Look at this. It’s straight from the official guide.

Your agent skips get_customer and calls lookup_order using only the customer’s stated name, leading to misidentified accounts. What change would most effectively address this?

A) Add a programmatic prerequisite that blocks lookup_order and process_refund calls until get_customer has returned a verified customer ID

B) Enhance the system prompt to state that get_customer is mandatory before any order operations

C) Add few-shot examples showing the agent always calling get_customer first

D) Implement a routing classifier that analyzes each request and enables only appropriate tools

Stop. Which one?

The answer is A. Here’s why.

When critical business logic has financial consequences, programmatic enforcement gives you guarantees. B and C rely on probabilistic LLM compliance. They’ll work 95% of the time and fail the other 5% in ways that cost you real money. D addresses tool availability, not tool ordering.

If you remember one rule for the whole exam, it’s this. When something has to work 100% of the time, don’t trust a prompt.

Hooks, prerequisites, deterministic code. Those are the senior moves.

Subscribe now

The fastest study path (this is what I’d run)

Here’s the playbook.

Step 1. Complete all 13 free Anthropic Academy courses.

Go to anthropic.skilljar.com. Create a free account. They have:

• Claude 101: Foundations

• AI Fluency

• Claude Code

• Model Context Protocol

• Claude API

Plus eight more. All free. All built by Anthropic engineers. Certificates auto-post to LinkedIn. Total time is about 10 hours.

Step 2. Ship one real project using Claude Code and MCP.

Doesn’t matter what. Build a research agent, a customer support bot, a code review tool. What matters is that it uses agentic loops, subagents, and at least one MCP server. The exam will find out if you’ve actually built this stuff.

Step 3. Take DeepLearning.AI’s Agentic AI course plus AI Agents in LangGraph.

These two hammer home the four agent patterns (Reflection, Tool Use, Planning, Multi-Agent) that show up everywhere on the exam.

Step 4. Register through Claude Partner Network and book the exam.

Free to join. Free for the first 5,000 partner network employees. $99 per attempt after that.

Step 5. Drill the 12 official sample questions and the 6 scenarios.

I built a full study guide with all 12 sample questions, domain breakdowns, and anti-patterns to avoid. If you want it, reply to this email and I’ll send it over.

Download the guide here.

The meta-lesson from going through all this

The biggest takeaway for me wasn’t a specific concept.

It was this. AI work in 2026 is shifting from “can you prompt it to do the thing” to “can you architect a system where the thing is guaranteed to happen correctly.”

Prompting alone is the junior move. Hooks, prerequisites, structured outputs, multi-agent coordination, context management. That’s the senior move.

And if you’re trying to break into AI engineering or level up this year, studying for this cert will force you to think like the latter, not the former.

Even if you never sit the exam, the study path alone is worth your weekend.

If you are finding value in my AI action letters, please consider subscribing. It keeps me going.

If you know someone trying to break into AI engineering or sitting on the fence about getting certified, please share this with them. This is the first cert I’ve seen that actually means something.

That’s it from me today.

Till next time. Stay tuned as I will share best resources from both my Harvard and Google networks to bring the best to you. Let’s up skill together.

Aspyre higher!

Abhijay

A friend of mine graduated with a CS degree last May.

She spent three months applying to jobs. Sent out 200+ applications. Got 4 callbacks.

Then one week, she got a referral at Google from someone she went to school with.

And a referral at Meta from a former intern she’d met once at a hackathon.

Both turned into interviews. One turned into an offer.

She didn’t suddenly become a better candidate. She just finally asked the right people.

Here’s the uncomfortable truth about job hunting — and honestly, about any kind of outreach:

Most people have the connections they need. They just don’t know it.

Your LinkedIn network is full of people at your dream companies. People who could refer you. People who could introduce you to a hiring manager. People who replied to your message two years ago and would happily help you today.

But you don’t know who they are. Because finding them means scrolling through hundreds of profiles, guessing at titles, wondering if they even check LinkedIn anymore.

So you do what everyone does: you apply cold. You send the same generic message to 10 people and hope someone replies.

Nobody does. And you wonder what you’re doing wrong.

I was frustrated by this. So I built something to fix it.

Subscribe now

Pledge a subscription. It means you value my content, and encourages me to write more.

Introducing your personalized LinkedIn CRM

The Insight: Your LinkedIn Export Is a Goldmine You’ve Never Opened

LinkedIn lets you download everything. Your full connection list. Your entire message history. Every connection request ever sent or received.

It’s sitting there, completely free, waiting to be used.

Go to: LinkedIn → Settings & Privacy → Data Privacy → Get a copy of your data

Select “Connections”, “Messages”, and “Invitations.” You’ll get a download link within 24 hours.

What you get:

• Connections.csv — every person you’re connected with, their title, company, and exactly when you connected

• messages.csv — your full DM history. Every conversation, every reply, every message you sent that got ignored

• Invitations.csv — every connection request sent and received. This one is underrated.

That last file tells you something most people miss.

If someone sent YOU a connection request, they found you organically. They were scrolling LinkedIn, saw your profile, and decided you were worth connecting with. That means they’re actively using the platform right now. That’s a signal you can use.

What I Built: A Personal CRM Engine Powered by Claude Code

Using Claude Code, I built a skill — essentially a set of instructions plus Python code — that takes this data export and turns it into a full CRM engine.

You give it two things: your LinkedIn data dump folder, and your goal.

It gives you back: the top 5 people to message, ranked by likelihood of success, with a personalized draft DM for each one.

Here’s how it works under the hood.

Step 1: Score Every Connection Across 5 Dimensions

Every single one of your connections gets a score out of 100. The score answers one question: given my specific goal, how much is this person worth reaching out to right now?

The five dimensions:

Relevance — 30 points

How closely does their title and company match what you’re trying to do? If you’re looking for a referral at Stripe and they’re a PM at Stripe, that’s full points. If they used to work at a fintech company, that’s partial credit.

The skill expands your search terms automatically. Looking for a “brand deal”? It also searches for marketing, partnerships, growth, devrel, sponsorships — because people don’t always use the same words you do.

Seniority — 20 points

Decision-making power matters. A CEO or Founder scores 20. A VP scores 15. A Director scores 10. A Manager scores 5. An intern scores 0 for seniority — not because they can’t help, but because they can’t make decisions.

For job hunting specifically, you want the hiring manager or someone senior enough to submit a referral that actually gets looked at. Not just anyone at the company.

Relationship Warmth — 25 points

This is the most important dimension. And it’s one only your message history can answer.

• Did they reply to a message you sent? 25 points. They’re responsive. They know who you are.

• Did you message them and get silence? 5 points. They exist but you haven’t broken through.

• Never messaged at all? 15 points. Clean slate — no baggage, fresh outreach.

The insight here: a cold contact is often better than a ghosted one. If someone ignored your message twice, their score drops. The skill knows not to send you back to a dead end.

LinkedIn Activity — 15 points

Are they actually using LinkedIn right now? The skill checks:

• Did they reply to any of your past messages? (+10) — confirmed active user

• Did they send YOU a connection request? (+8) — they’re actively browsing

• Did you connect recently? (+5 for last 30 days, +3 for last 90)

• Do you have their email from the export? (+2) — option for multi-channel

Recency — 10 points

Someone you connected with last month is more likely to remember you than someone from 2019. Recency of connection is a proxy for how top-of-mind you are.

Step 2: Find Who Can Bridge You to People You Can’t Reach Directly

Here’s where it gets interesting.

LinkedIn’s export doesn’t include your 2nd degree connections. You can’t see who your connections know.

But you can infer it.

The skill uses company clustering: for any target company, it finds every 1st degree connection who currently works there. Those people are your bridges — even if they’re not in the exact team or role you need, they’re inside the building.

Say you want a referral at Google and you have 40 Google connections. The skill doesn’t just pick the most senior one. It picks the one who:

• Is in the right org (engineering if you’re an engineer, product if you’re a PM)

• Has actually replied to your messages before

• Connected with you recently

• Has a seniority level high enough to submit a meaningful referral

It also finds super connectors — people in your network with BD, partnerships, or investor titles who are naturally well-connected across companies. These people often know exactly who you need to reach, even at companies where you have zero connections.

Step 3: Draft Messages That Don’t Sound Like Every Other LinkedIn DM

This is the part that required the most thought.

Generic messages get ignored. “Hey, hope you’re doing well! I’d love to pick your brain…” No. Stop.

The skill classifies every contact into a role type — marketing, founder/exec, technical, investor, or other — and drafts a different message for each.

Here’s how it branches:

If they’re in marketing or partnerships: Lead with your audience size and a specific brand partnership angle. They care about reach and fit.

If they’re a founder or C-suite: Lead with “love what you’re building at [Company].” They’re proud of what they’ve built. Acknowledge it.

If they’re an engineer or technical role: Don’t pitch them directly. Ask them to redirect you. “Any chance you could point me to someone on your partnerships team?” Engineers don’t make sponsorship decisions — but they can open a door.

If they’ve replied to you before: Never open with a fake warm line. Open with a real one: “Following up on our last exchange.” Only when they actually replied.

If you messaged them before and got silence: “Circling back —” is your opener. Short. Direct. No pretending it’s a fresh message.

The result: every message sounds like it was written specifically for that person. Because it was.

Step 4: Verify They’re Still There Before You Send Anything

LinkedIn data exports are snapshots. People change jobs. The export still shows their old company.

I learned this the hard way when I ran the skill and my top-ranked contact at a company turned out to have left months ago.

So the skill now runs a quick web search on the top 3 contacts before drafting their messages. If someone’s moved on, they get flagged and dropped. The next-ranked person moves up.

Stale data is the silent killer of LinkedIn outreach. Most people don’t check. They send a message about “your work at [Company]” to someone who left a year ago. It’s immediately obvious you didn’t do your homework.

What This Looked Like in Practice

When I ran the skill for a job referral goal at a target company, it found:

• 3 warm two-way contacts at the company I’d forgotten I’d ever spoken to

• 1 super connector — a VC in my network with portfolio companies at the target — who could make a warm intro to the founder directly

• 8 contacts to re-engage — people I’d messaged when I had a much smaller following who never replied. The message was stale. The relationship wasn’t dead. New message, new stats, new chance.

The biggest win wasn’t the top-ranked contact. It was realizing I had a warm path I’d completely overlooked.

How to Run This Yourself

You need three things:

1. Your LinkedIn data export — request it at LinkedIn Settings → Data Privacy → Get a copy of your data. Takes 24 hours.

2. Claude Code — Anthropic’s CLI tool. Free tier works fine for this.

3. The skill file — I’m publishing the full .md file on GitHub [LINK COMING SOON]. It’s a single file. Drop it in your Claude skills folder and you’re done.

Then run:

/linkedin-outreach [your specific goal here]

Be specific. “Find me a job” won’t work. “Find me someone at Stripe’s PM team who can refer me for an APM role” will give you a ranked list with drafted messages in under 60 seconds.

The Mindset Shift

Most people treat networking like a lottery. Send a bunch of messages, hope someone responds, repeat.

This turns it into a system.

You know exactly who to message. You know exactly what to say. You know who’s warm, who’s cold, who’s active, and who can bridge you where you need to go.

Your network hasn’t changed. Your ability to read it has.

That friend who got the Google and Meta interviews? She had the connections the whole time. She just needed someone to tell her who to ask.

Now you have a tool that does exactly that.

If you are finding value in my AI action letters, please consider subscribing. It keeps me going.

The full skill file is live on GitHub: github.com/abhijayarora93/linkedin-crm-skill — setup instructions are in the README. Read here. And I’ll be dropping a walkthrough video in the next issue, so make sure you’re subscribed.

If you know someone who’s been grinding through job applications and not hearing back, please share this with them. This is exactly the edge they need.

That’s it from me today. Subscribe please before leaving and pledge a subscription! :)

Till next time. Stay tuned as I will share best resources from both my Harvard and Google networks to bring the best to you. Let’s up skill together. Aspyre higher!

Hey folks,

Last week a friend from my Harvard network sent me a message that I’ve now gotten dozens of times:

“I’m a Product Manager. I want to become an AI Product Manager. Where do I even start?”

The honest answer? Most career transition advice is garbage. “Learn Python.” “Take a course.” “Network more.” Thanks, super helpful.

Here are a few resources I recommend:

For someone starting from zero and wanting to actually understand AI (not just play with ChatGPT), here are the three pieces I’d recommend, in this exact order:

1. 3Blue1Brown’s “Neural Networks” series on YouTube → Full playlist here

Start with these two essentials:

• “But what is a Neural Network?” (Chapter 1)

• “Attention in transformers” (the chapter on how ChatGPT actually works)

Grant Sanderson’s animated explainers are the best foundation you can build. Free, visual, no fluff. After this, you’ll never feel lost when people throw jargon around.

2. Andrej Karpathy’s “Intro to Large Language Models” → Watch here (1 hour)

Karpathy is ex-OpenAI, ex-Tesla AI, and explains LLMs the way a senior practitioner explains it to a smart friend. This bridges the gap between “I get neural networks” and “I get how ChatGPT actually works, what its limitations are, and where the field is going.” If you only watch one thing, watch this.

Want to go deeper? His longer “Deep Dive into LLMs” (~3 hours) is the natural next step.

3. Anthropic’s “Building effective agents” + a DeepLearning.AI short course → Anthropic guide + DeepLearning.AI courses

Now you go applied. Anthropic’s guide teaches you how AI is actually being deployed in products today. Pair it with one of Andrew Ng’s free short courses — start with ChatGPT Prompt Engineering for Developers or AI Prompting for Everyone if you’re less technical. This is where you go from understanding AI to building with it.

The sequence matters: fundamentals → how LLMs work → how to build with them. Most people skip steps 1 and 2 and jump straight to “10 ChatGPT prompts that will change your life” content, which is why they stay confused. These three will take you maybe 6-8 hours total and you’ll be ahead of 95% of people talking about AI on LinkedIn.

But I also built something different. A single prompt that scrapes real AI job postings, analyzes YOUR actual resume, identifies YOUR specific skill gaps, and generates a personalized 90-day roadmap — with free courses, weekly deliverables, and resume rewrites.

No generic advice. No $2,000 bootcamp. Just a brutally honest, hyper-specific plan built from what companies are actually hiring for right now.

Here’s exactly how it works — and the full prompt you can copy-paste today.

The problem with “just upskill”

When you Google “how to transition to AI Product Manager,” you get the same recycled advice: learn machine learning basics, get a certification, update your LinkedIn headline.

But here’s what nobody tells you: the skills companies actually list in AI PM job descriptions change every few months. Six months ago, everyone wanted “experience with LLMs.” Now? It’s “evaluation frameworks,” “responsible AI governance,” and “prompt engineering for production systems.”

If you’re studying from a static curriculum, you’re already behind.

The move is to reverse-engineer what companies are hiring for right now — this week — and build your learning plan around that.

How this prompt works (3 steps)

1. It scrapes real jobs posted in the last 7 days

The prompt uses Apify’s LinkedIn Jobs Scraper to fetch the 15 most recent AI versions of whatever your current role is — posted in the past week, in the US. These aren’t hypothetical job descriptions. They’re live postings from companies actively hiring.

From those 15 JDs, it extracts the top recurring skills — the ones that show up again and again across different companies. That’s your signal. That’s what the market actually wants.

2. It cross-references with YOUR resume

This is where it gets personal. The prompt takes your actual resume text and compares it line-by-line against those recurring skills. It identifies your top 5 skill gaps — and it’s specific.

Not “you need more AI experience.” More like: “You have strong stakeholder management from your 3 years at Google, but zero evidence of ML model evaluation, prompt engineering for production systems, or responsible AI framework knowledge. Your consumer product launches are impressive but don’t demonstrate any AI-native product thinking.”

Brutal? Yes. Useful? Extremely.

Subscribe now

3. It builds a 90-day interactive roadmap

Here’s the part that blew my mind. The prompt generates a full Notion-style interactive HTML dashboard with:

• A progress bar and stat cards that update in real time as you check off courses

• Week-by-week course lists (Weeks 1-12) organized under collapsible Month 1/2/3 sections

• Every course is free — Coursera audit tracks, DeepLearning.AI, fast.ai, Google certs, Microsoft Learn, HuggingFace tutorials

• Color-coded skill gap tags so you can see which gap each course addresses

• “Open” buttons that link directly to each course URL

• Weekly deliverables — not just “watch videos” but actual outputs like “build an AI product spec” or “rewrite your resume with AI terminology”

• Resume and LinkedIn rewrites tailored to the gaps you’re filling

• A target company list pulled from the actual job postings it scraped

33 items across 12 weeks. All interactive. All personalized to YOUR resume.

Here is my output:

The full prompt (copy-paste this)

You’ll need two things before you run this:

• Your resume text (copy-paste from your PDF)

• An Apify account (free tier works) with the curious_coder/linkedin-jobs-scraper actor

Subscribe now

Here’s the prompt:

You are an AI Career Strategist and expert front-end developer.

I am a [CURRENT ROLE] at [COMPANY] with [CREDENTIALS].
My resume is below. I want to transition to AI [CURRENT ROLE].

[PASTE YOUR RESUME HERE]

Do all of the following in one response:

1. Scrape jobs
Use Apify (curious_coder/linkedin-jobs-scraper) to fetch 15 real
AI [CURRENT ROLE] jobs posted in the past 7 days in the US.
Extract the top recurring skills across all JDs.

2. Analyze my gaps
Cross-reference those skills against my resume. Identify my top 5
skill gaps. Be brutally honest and hyper-specific to MY actual
resume — name my real employers, real skills, real gaps.
No generic advice.

3. Build a Notion-style interactive dashboard
Build a single self-contained HTML file with:

- Properties block: Status, Target role, Target companies,
  Salary range, Courses completed counter, Skill gaps
  (color-coded pills), Your leverage
- Live progress bar + 4 stat cards (total hours, % free,
  done, remaining) that update in real time
- Brutally honest callout box specific to my resume
- Week-by-week course list (Weeks 1-12) under collapsible
  Month 1/2/3 sections, each week also collapsible
- Each course row: checkbox, title, color-coded skill tag,
  source, hours, and an Open button linking to the real
  free course URL
- Mini dot indicators on each week header showing completion
- Each week ends with a Deliverable row
- Resume + LinkedIn rewrites based on gaps
- Target company list from the scraped postings
- Start over button to reset

Only use FREE resources: Coursera free audit, YouTube,
DeepLearning.AI, fast.ai, Google, Microsoft Learn, HuggingFace.

Replace [CURRENT ROLE], [COMPANY], [CREDENTIALS], and paste your actual resume where indicated.

How to run it

Option 1: Claude (what I used)

Paste the prompt directly into Claude. It will generate the full HTML dashboard in one shot. Copy the HTML output, save it as a .html file, and open it in your browser. You’ve got a fully interactive career transition dashboard.

Option 2: Build the automated version

If you want to get fancy, you can wire this into an n8n workflow:

• Apify node → scrapes jobs on a schedule

• Claude/GPT node → runs the analysis and generates the dashboard

• Gmail node → emails you the updated dashboard weekly

This way your roadmap stays current as job requirements shift.

Why this actually works

I’ve seen too many people spend 6 months “learning AI” with no direction, then realize they studied the wrong things. This approach flips it:

You start with what companies want today. You end with a plan that’s specific to what YOU are missing.

The color-coded tags are key. When you see that 4 of your 5 gaps are in “LLM Depth” and “Evaluation Frameworks,” you know exactly where to spend your limited time. No guessing.

And because every course is free, the only cost is your time. No excuses.

One concrete action for today

Copy the prompt above. Paste your resume in. Run it in Claude. Save the HTML file and open it.

You’ll have your personalized 90-day plan in under 5 minutes. Then check off your first course this week.

The people breaking into AI roles right now aren’t smarter than you. They just have a system. Now you do too.

Subscribe now

If you are finding value in my AI action letters, please consider subscribing. It keeps me going.

If you know someone trying to break into an AI role — PM, engineer, designer, analyst, doesn’t matter — please share this with them. The prompt works for any job title.

That’s it from me today.

Till next time. Stay tuned as I will share best resources from both my Harvard and Google networks to bring the best to you. Let’s up skill together. Aspyre higher!

Today, something happened that stopped me mid-scroll.

OpenAI shut down Sora. Their AI video generator. The one that topped the App Store. The one Disney was going to invest $1 BILLION into.

Gone. Six months. That’s all it lasted.

The reason? The costs were too high, and not enough people used it consistently.

And on the exact same day, Anthropic quietly shipped a feature called Dispatch — and honestly, it’s the most useful AI update I’ve seen all year.

Here’s why I think this moment matters more than people realize.

The AI industry has a “cool demo” problem.

Everyone’s racing to build the flashiest thing. Video generation. Image creation. Voice cloning.

But almost nobody is building AI that fits into how you actually work every day.

That’s exactly what Dispatch does differently.

So What Is Dispatch?

Dispatch is a new feature inside Claude (Anthropic’s AI) that lets you text Claude a task from your phone — and it executes on your Mac in the background.

Read that again.

You’re on the subway. You text: “Prep me for my 2pm meeting with Sarah from Stripe.”

By the time you get to the office, Claude has researched Sarah’s LinkedIn, pulled up recent Stripe news, generated 5 talking points, and drafted 3 smart questions to ask.

All sitting on your desktop, ready to go.

Here’s what makes this powerful:

1. It uses YOUR computer, YOUR files, YOUR apps

This isn’t some cloud chatbot. Claude literally opens apps on your Mac, clicks through browsers, fills spreadsheets, writes code. It has access to your local files, your project folders, everything on your machine.

2. It’s asynchronous

You assign the task and walk away. Go to the gym. Cook dinner. Sleep. Come back to finished work.

3. It gets smarter when you build custom skills

This is the part most people are missing — and the part I want to spend the rest of this article on.

Subscribe now

What Are Claude Code Skills?

Claude Code is Anthropic’s CLI tool for power users. You run it in your terminal.

But here’s the unlock: you can create custom skills as simple .md (markdown) files.

Each skill is a set of instructions that tells Claude exactly how to do a specific task — in YOUR voice, with YOUR preferences, targeting YOUR files.

Think of it like creating a recipe card for your AI assistant. Write it once, and Claude follows it perfectly every time.

You save these files in a folder called ~/.claude/skills/ and Claude reads them automatically.

No coding required. Just plain English instructions in a text file.

10 Skills I Built (And You Can Too)

I spent the last few weeks building custom skills for the tasks that eat up my time. Here are the 10 that have made the biggest difference.

Every single template below is copy-pasteable. Save it as a .md file in your ~/.claude/skills/ folder and you’re good to go.

1. Email Manager

This one scans my entire inbox, categorizes every email by priority (Urgent / Important / FYI / Skip), checks for unanswered follow-ups, and drafts replies in my writing style.

The key detail: I wrote my exact email style into the skill file. Short. Casual greeting. No filler phrases. Direct pricing for brand inquiries.

So the drafts actually sound like me — not like ChatGPT.

With Dispatch: I text “check my email” while brushing my teeth. By the time I sit down with coffee, there’s a prioritized summary with draft replies waiting for my approval.

Save as email-manager.md:

# Email Manager

Scan my inbox, categorize emails, and draft replies in my style.

## Instructions

### Step 1: Read Inbox
Read emails from the last 24 hours.
Extract: sender, subject, date, read status, body (first 2000 chars).

### Step 2: Categorize
- URGENT — action needed today (deadlines, money, direct asks)
- IMPORTANT — reply within 48h (opportunities, key contacts)
- FYI — no reply needed (newsletters, notifications)
- SKIP — auto-skip (no-reply addresses, promotions)

### Step 3: Check Follow-ups
Scan Sent Mail for emails I sent in the last 7 days with no reply.

### Step 4: Generate Summary
Output grouped by tier with one-line context for each.

### Step 5: Draft Replies
For URGENT and IMPORTANT, draft in my style:
- Brief (1-3 sentences), casual greeting, no filler phrases
- NEVER send without my explicit approval

2. Competitor Research

This skill monitors competitors across the web — product launches, pricing changes, hiring patterns, press coverage — and compiles a one-page intelligence brief.

It searches TechCrunch, LinkedIn, Product Hunt, company blogs, and Twitter in parallel.

With Dispatch: Sunday evening text: “What did [Competitor] ship this week?” Start Monday’s standup with fresh intel nobody else has.

Save as competitor-research.md:

# Competitor Research

Research competitors and deliver a structured intelligence brief.

## My Competitors
- [Company A] — [website]
- [Company B] — [website]

## Instructions

### Step 1: Search for Recent Activity
For each competitor, search:
- "[company] product launch OR announcement 2026"
- "[company] pricing change OR new feature"
- "[company] hiring site:linkedin.com"
- "site:techcrunch.com [company]"

### Step 2: Analyze
Extract: new features, pricing changes, key hires, press coverage, sentiment.

### Step 3: Output Brief
## Competitor Brief — [Date]
### [Company]
What's New: [1-2 sentences]
Signal: [What this means strategically]
Threat Level: Low / Medium / High

### Key Takeaways
- [3 bullets: what matters most for us]

3. Meeting Prep

Give it a meeting topic and attendee names. It researches each person’s background, recent activity, company news, and generates talking points + smart questions to ask.

I never walk into a meeting unprepared anymore.

With Dispatch: Text from the Uber on the way. Arrive with a full dossier.

Save as meeting-prep.md:

# Meeting Prep

Research attendees and prepare a meeting brief with talking points.

## Instructions

### Step 1: Research Each Attendee
Search: "[name] [company] site:linkedin.com/in"
and "[name] interview OR keynote 2025 2026"
Extract: role, background, recent activity, known opinions.

### Step 2: Research the Topic
Find recent developments, data points, trends.

### Step 3: Generate Brief
## Meeting Brief: [Topic]
### Attendee Profiles
[Name] — [Title] at [Company]
- Background, recent activity, conversation starter

### Talking Points (5 prepared points with data)
### Smart Questions to Ask (3 questions showing homework)

4. Social Media Writer

This one scrapes top-performing posts on any topic, analyzes what hooks and formats drove engagement, then rewrites in my exact voice.

I baked my writing style rules into the file — line breaks, tone, CTA patterns, things I never say. The output is ready to paste.

With Dispatch: See breaking news on your phone? Text Claude immediately. By the time you open your laptop, there’s a researched, formatted post ready to publish while the topic is still trending.

Save as social-writer.md:

# Social Media Writer

Write social media posts in my voice by researching what's working.

## My Writing Style
- Tone: [conversational / professional / witty]
- Format: [one thought per line / short paragraphs]
- Hooks: [bold statements / questions / contrarian takes]
- CTAs: [comment keyword / repost / link in comments]
- NEVER do: [emojis everywhere / corporate speak]

## Instructions

### Step 1: Research viral posts on the topic
### Step 2: Analyze top performers (hook, format, trigger, CTA)
### Step 3: Rewrite in my voice, under 1300 characters
Output in a code block for easy copy-paste.

Subscribe now

5. Weekly Report Generator

Reads my project files, notes, and task lists. Compiles a structured weekly report — accomplishments, metrics, blockers, next week’s priorities.

Same clean format every week. My manager loves the consistency.

With Dispatch: Friday afternoon text: “Generate my weekly report.” Review and send before EOD without opening a single spreadsheet.

Save as weekly-report.md:

# Weekly Report Generator

Compile a weekly status report from my project files.

## My Project Files
- Notes: ~/Documents/work-notes/
- Metrics: ~/Documents/metrics.csv
- Tasks: ~/Documents/todo.md

## Report Format
# Weekly Report — [Date Range]
## Accomplishments (from completed tasks and recent changes)
## Key Metrics (compare to last week)
## Blockers (stalled items or overdue tasks)
## Next Week's Priorities (from upcoming tasks)

6. File Organizer

Scans my Downloads folder (which is always a disaster), categorizes files by type, moves them into organized subdirectories, and renames them with consistent conventions.

It shows a preview of all changes before executing — never moves anything without approval.

With Dispatch: Text “organize my Downloads folder” once a week. 200+ files sorted while you’re making dinner.

Save as file-organizer.md:

# File Organizer

Organize a target folder by categorizing and moving files.

## Target Folder
~/Downloads/

## Rules
- PDFs → ~/Documents/PDFs/[YYYY-MM]/
- Images (jpg, png, gif) → ~/Pictures/Sorted/[YYYY-MM]/
- Videos (mp4, mov) → ~/Videos/Sorted/
- Code (py, js, html) → ~/Documents/Code/
- Spreadsheets (csv, xlsx) → ~/Documents/Spreadsheets/

## Instructions
1. Scan folder and list all files
2. Categorize each by extension
3. Show preview: [current name] → [new location]
4. Ask for approval before moving anything
5. Output summary with counts per category
6. Flag duplicates (same name + same size)

7. Birthday & Date Reminders

Reads a contacts CSV with birthdays and important dates. Checks what’s coming up in the next 7 days. Drafts personalized messages based on my relationship with each person.

No more “oh no, it was yesterday” moments.

With Dispatch: Monday morning text: “Check birthdays this week.” Warm, personalized messages drafted and ready to send.

Save as birthday-reminder.md:

# Birthday & Date Reminder

Check upcoming birthdays and important dates, draft messages.

## Data Source
~/Documents/contacts-birthdays.csv
(Format: Name, Birthday MM/DD, Relationship, Notes)

## Instructions
1. Find dates in the next 7 days
2. Draft a message for each:
   - Close friends: warm, personal, reference something specific
   - Professional: friendly but brief
   - Acquaintances: simple and genuine, 1-2 sentences
3. Present messages for approval
4. Show "Coming Up" for dates 8-30 days away

8. Resume Tailor

Takes my master resume and a job description URL. Produces a tailored resume that mirrors the job’s keywords, reorders experience by relevance, and optimizes for ATS systems.

It never fabricates experience — just reframes what’s real.

With Dispatch: Found a dream job on your phone? Text Claude the URL. Tailored resume ready by the time you get to your laptop.

Save as resume-tailor.md:

# Resume Tailor

Customize my master resume for a specific job description.

## My Master Resume
~/Documents/resume-master.md

## Instructions
1. Read my master resume
2. Fetch the job description URL — extract required skills,
   qualifications, responsibilities, technologies
3. Tailor the resume:
   - Reorder sections by relevance to this role
   - Rewrite bullets using their keywords
   - Add tailored Summary section
   - Ensure ATS compatibility (no tables, standard headers)
4. Save as ~/Documents/resumes/Resume_[Company]_[Role].md
5. Show a Match Score (% of requirements covered)

RULE: NEVER fabricate experience. Only reframe what's real.

9. Morning News Briefing

Searches the web for top stories in my chosen topics (AI, tech, markets), filters out noise, and delivers a personalized 2-minute briefing with a “water cooler” story for small talk.

With Dispatch: Text “morning briefing” when your alarm goes off. Know everything important before your first meeting — while you’re still in the shower.

Save as morning-briefing.md:

# Morning News Briefing

Deliver a personalized daily news digest.

## My Topics
- Artificial Intelligence (models, launches, policy)
- Tech industry (big tech, startups, funding)
- [Your Industry] (add your sector)
- Markets (S&P 500, notable movers)

## Output Format
# Morning Briefing — [Date]
## The Big Story
[Headline + 3-sentence summary + why it matters]

## Quick Hits
- [Topic]: [1-sentence summary] — [source]
(repeat for 3-5 stories)

## Water Cooler
[1 fun/interesting story for small talk]

## Deep Dives (if you have time)
- [Link]: [why worth reading]

10. Expense Tracker

Reads receipt photos (Claude can see images), categorizes expenses, tracks against monthly budget limits, and generates spending reports.

With Dispatch: Snap a receipt at a restaurant, text it to Claude. It extracts the amount, categorizes it, updates your tracker. End of month? “Generate my expense report.” Done.

Save as expense-tracker.md:

# Expense Tracker

Track expenses, categorize spending, generate reports.

## Data File
~/Documents/expenses-2026.csv
(Format: Date, Amount, Category, Description, Source)

## Budget Limits (monthly)
- Food & Dining: $800
- Transport: $300
- Subscriptions: $200
- Shopping: $500
- Entertainment: $200

## When logging a new expense:
1. If given a receipt image — read it, extract merchant + amount + date
2. Categorize based on merchant name
3. Append to CSV
4. Show running total for that category vs budget

## When generating a report:
1. Sum by category for the month
2. Compare to budget limits
3. Flag over-budget categories
4. Show top 5 largest expenses
5. Calculate total spend and savings rate

Why This Matters More Than Sora

Here’s my honest take.

OpenAI spent massive resources building Sora — a video generator that looked incredible in demos but didn’t stick in people’s daily lives. Six months and it’s gone. Disney walked away from a billion-dollar investment.

Meanwhile, Anthropic built something that costs a fraction of the compute and solves real problems people have every single day.

The lesson: The AI products that win won’t be the flashiest ones. They’ll be the ones that disappear into your workflow so seamlessly you can’t imagine going back.

That’s what custom skills + Dispatch does. It’s not a demo. It’s infrastructure for how you work.

How to Set This Up From Scratch

You don’t need to know how to code. If you can write a bullet-point list, you can build a Claude Code skill.

Here’s the full setup in 5 steps:

Step 1: Install Claude Code

You need Node.js first. If you don’t have it, open Terminal and run:

brew install node

Then install Claude Code:

npm install -g @anthropic-ai/claude-code

Step 2: Log in

Run claude in your terminal. It’ll open a browser to authenticate with your Anthropic account. You need a Pro ($20/mo) or Max ($100/mo) plan.

Step 3: Create your skills folder

mkdir -p ~/.claude/skills/

Step 4: Add your first skill

Copy any template from this article into a .md file:

nano ~/.claude/skills/email-manager.md

Paste the template, customize the parts in brackets [like this] with your own details, and save.

Step 5: Set up Dispatch

Install the Claude Desktop app on your Mac and the Claude mobile app on your phone. Open Claude Desktop, go to the Dispatch section, click “Get Started”, configure permissions, and scan the QR code with your phone to pair. Start texting tasks from anywhere.

Pro Tips for Better Skills

Be specific about YOUR style. The more detail you put in the “my style” section, the less you’ll edit outputs. Include examples of your past work.

Include file paths. Tell Claude exactly where your files live. ~/Documents/metrics.csv beats “my metrics file.”

Add safety rails. For anything that sends emails or moves files, always include “show me a preview before executing.”

Iterate. Your first version won’t be perfect. Run the skill, note what you’d change, update the .md file. After 2-3 rounds, it’ll feel like magic.

Combine skills. The real power is chaining: “Check my email, then prep me for any meetings mentioned in urgent emails.”

If you are finding value in my AI action letters, please consider subscribing. It keeps me going.

If you know someone who spends too much time on repetitive tasks — organizing files, prepping for meetings, writing reports, managing email — please share this with them. This could save them 10+ hours a week.

That’s it from me today.

Till next time. Stay tuned as I will share best resources from both my Harvard and Google networks to bring the best to you. Let’s up skill together. Aspyre higher!

Shortscast is my new initiative where I interview CXOs and founders in Silicon Valley and bring top 1% of insights in 60s Reels/Shorts format here.

Jeff has been at Replit for four years. He was there when they did a 30-40% layoff and got down to 50 people. He stayed because he believed in the mission: empower a billion people to build software.

Now Replit is ripping — hundreds of enterprise customers, Agent 4 about to drop, and a thesis that’s proving itself every single day.

Here is the photo clicked, 4 days before the launch of Agent 4 at Replit’s office!

And here’s what stood out from our conversation.

1. A real estate firm in Ohio built their app for $100

Jeff told me about a real estate firm in Ohio. They needed an app. Before vibe coding, they would have had to pay a developer $10,000/month on retainer. Completely unaffordable for a small business.

Instead, they built it on Replit for $100.

“We have somebody monetizing ice skating rinks in the UK. We have hundreds of these examples. They would never be able to afford a developer. You can now just build that app yourself, and it will change your business.”

This is what excites me most about vibe coding. It’s not the billion-dollar one-person company that gets all the LinkedIn hype. It’s the small businesses that previously couldn't exist that can now.

A niche local business in rural America built an online storefront on Replit that became a clear differentiator from its peers. For $100. That’s the real revolution.

2. Your first working version used to take 6 months. Now it takes hours.

On the enterprise side, Jeff broke it down into two buckets: prototyping and internal tools.

For prototyping, the impact is insane:

“Instead of bullets that you go back and forth with engineering for months at a time, you just make that, you test it with the user, and you have a really good idea of what needs to be built. We’ve had customers say your first working version takes six months — now you can do it in hours.”

Think about what that means. A PM can validate an idea before ever taking an engineer’s time. Instead of 100 non-validated ideas going to engineering, you come with the 2-3 best, most polished ideas already tested with users.

For internal tools, it’s even simpler. If you’re anywhere in a large company and need a dashboard, an automation, or something to move faster — you just build it yourself. No tickets. No waiting 6 weeks for IT.

Subscribe now

3. Everyone is becoming a generalist. The roles are blurring.

I asked Jeff about how product, engineering, and design roles are starting to overlap. His answer was clear:

“Our general thesis is that everyone becomes more of a generalist and you’re able to self-serve in a lot more ways.”

This doesn’t mean specialists disappear. Jeff still hires people who peak in one area. But he wants them to be generalist enough that if they need to self-serve something — build a prototype, create an internal tool, vibe code a slide deck — they just do it without waiting for someone else.

The days of “that’s not my job” are numbered.

4. Someone applied to Replit with an app instead of a resume

This one blew my mind.

“One of the people currently on our team built — instead of a resume — a Replit app. It takes two minutes. But they included a quiz at the bottom to see if I actually read it. Those are the type of people I want to hire.”

Jeff confirmed: everyone applying to Replit should know vibe coding.

It’s so easy now to produce something novel and creative in minutes that if you’re still sending a PDF resume, you’re already behind. The bar has shifted from “can you describe what you’ve done” to “can you show me what you can build?”

If you’re job hunting right now, this is your sign. Build something. Ship it. Link it.

5. They survived a 30-40% layoff. The people who stayed cared about the mission.

Jeff has been at Replit for four years. The first two and a half? The business was not going well.

“We did a 30-40% layoff and got down to 50-60 people. The people who stayed when it was awful were the people who cared about what we were doing. That is why I stayed — from the first day I joined, it was to empower people to build software.”

At the time, they didn’t know AI would be the thing that accelerated them. But the people who stayed through the worst times are the ones doing the best work now. Jeff still filters for that in hiring — because hard times will come again, and you want people who stay because they believe, not because things are going well.

This is an underrated career lesson. Join the mission, not the hype cycle.

6. Agent 4: It’s like having a team working for you

Jeff gave me a preview of what’s coming with Replit’s Agent 4:

Parallelization — multiple sub-agents doing work simultaneously. Instead of one agent doing one thing at a time, you’ll have a team of agents working in parallel. Jeff said it’s “a little intimidating at first because it’s a new muscle, but once you get the hang of it, it’s literally like having a team working for you.”

Beyond web apps — Replit already has video and animation stacks live. Slides are coming with a curated experience. The vision is clear: anything you want to do on a computer is code, and Replit wants to be the place where you run it.

7. A divorced parent built a co-parenting app in 2 hours

When I asked Jeff what AI use case blew his mind, he told me about a mobile app buildathon:

“Somebody built an app that makes it easier to co-parent for divorced parents — to help manage scheduling, expenses, things like that. It’s really cool that it didn’t exist for them. That’s a real world problem that somebody who is a non-developer finally was able to build for themselves.”

This is the thing that gets lost in all the “AI will replace everyone” discourse. AI is enabling people to solve their own problems. Problems so specific and personal that no developer would ever build a product for them. That’s powerful.

The bottom line

After talking to Jeff for 15 minutes, here’s what I took away:

- The vibe coding revolution isn’t about billion-dollar companies. It’s about the small business in rural Ohio that finally has a website.

- Your resume is not a PDF anymore. It’s what you can build in 2 minutes.

- Roles are blurring. If you can’t self-serve basic technical tasks, you’re going to fall behind.

- Join the mission, not the hype. The people who thrive are the ones who care about what they’re building, especially when times are tough.

- Agent 4 is going to surprise people. Parallelized sub-agents + extending beyond web apps = a much bigger Replit than most people realize.

Watch the full Shortscast episode with Jeff Burke here.

Of course, this is the AI Action Letter, and not just any other news letter so I had to build. I built a job hunting tool on Replit Agent 4, and the results blew my mind!

If you are new to Replit, you can use this link to get $10 in free credits.

Try the the app I built: https://job-hunt-agent.replit.app/

And here is how you can build one for yourself:

If you found this useful, subscribe to the AI Action Letter for more breakdowns like this every week. And if you know someone who’s thinking about vibe coding for their business — forward this to them. Let me know in the comments if you have any questions, and I will get to them. Until next time!

When I was job hunting, I would spend 3-5 hours applying to jobs.

Three. Hours.

I was tweaking my resume for each posting. Googling whether the company actually cared about TypeScript or just said they did. Wondering if my “5 years of experience” bullet was hitting the right keywords.

After all that effort? Maybe four applications.

I sat back and thought — I literally work with AI tools every single day. Why am I doing this manually?

So I built something.

I built a Job Hunt Agent that does the boring part of job hunting automatically. It scrapes LinkedIn for fresh postings (last 24 hours, your exact roles), scores every single one against your resume using Claude AI, and flags exactly which keywords you’re missing.

You go from a wall of jobs you have to individually evaluate to a ranked table where the best matches float to the top.

I built this on Replit. New to Replit? Use this link to get 1 month free (limited to 4 redemptions only). If this is expired, you can use this link to get $10 in free credits.

Here’s exactly how to set it up and use it yourself.

What the agent actually does

Before we get into the setup, here’s what’s happening under the hood:

• Scrapes LinkedIn for jobs posted in the last 24 hours matching your target roles and location

• Runs every job through an ATS scoring algorithm (Claude AI) that keyword-matches your resume to each listing

• Shows you a sortable, filterable table ranked by match score

• For any job you like, gives you a full optimization report: rewritten headline, rewritten summary, missing keywords, top 3 changes to make

You need two things to run it: an Apify token (free tier works) and your resume as a PDF.

That’s it. Let’s go.

Step 1: Get your Apify token

Apify is what does the LinkedIn scraping. Their free tier gives you enough credits to run dozens of searches.

→ Go to apify.com and create a free account → In your dashboard, click your profile → Settings → Integrations → Copy your Personal API token (starts with apify_api_...)

One concrete action: Grab your Apify token and keep it in a notes app — you’ll paste it in every time you run the agent.

Step 2: Enter your target roles and location

Open the Job Hunt Agent. In the Target Roles field, type the exact job titles you’re searching for, separated by commas.

→ Be specific: “AI Product Manager” outperforms “Product Manager” if that’s your niche → You can search multiple roles at once: “Software Engineer, Frontend Engineer” → Each role gets its own independent search — results don’t get mixed together → Set location to Remote, a city, or leave it blank for global results

One concrete action: Start with your two most-targeted job titles. You can always re-run with different roles — it only takes a few minutes.

Step 3: Upload your resume PDF

Drag and drop your resume PDF into the upload zone, or click to browse.

→ The resume is parsed entirely in your browser — it never leaves your device → The extracted text is what gets matched against every job description → Keep your resume up to date before running (obvious, but worth saying)

One concrete action: Use the resume you’d actually submit — not a rough draft. The ATS scores are only as useful as the document you feed it.

Step 4: Run the search and read your ATS scores

Hit Start Hunting. The agent will:

1. Kick off LinkedIn scrapes for each of your roles in parallel

2. Pull back postings from the last 24 hours

3. Auto-score every job against your resume (in batches so Claude doesn’t get overwhelmed)

The results table shows each job with a colored ATS score badge:

• Green (Strong, 72+) — Your resume already hits most of the keywords. Apply immediately.

• Yellow (Partial, 42–71) — Good fit but gaps exist. Worth a quick resume tweak.

• Red (Low, <42) — Poor keyword match. Skip or make major edits first.

You can sort by score, filter by experience level (Early / Mid / Senior), and set a minimum score threshold so you only see strong matches.

One concrete action: Sort by ATS score descending and focus on anything green first. Those are your highest-leverage applications today.

Step 5: Use the Optimize panel for any job you want to apply to

Click the Optimize Resume button on any job card. The agent fetches the full job description and runs a deeper analysis:

→ A breakdown of your skills match, experience match, and title match (0–100 each) → Your top 3 specific changes to make to your resume for this role → 5–8 keywords to weave into your bullets → A rewritten headline tailored to this job → A rewritten professional summary you can paste directly into your resume

Copy the rewritten sections into your resume, apply, repeat.

One concrete action: Run the Optimize panel on your top 3 green-scored jobs. You’ll have three tailored resume versions in under 10 minutes.

The prompts under the hood

I’m going to show you exactly what the AI is doing when it scores and optimizes. No black box.

ATS Scoring Prompt (runs automatically on every job):

You are an ATS (Applicant Tracking System) engine performing a strict keyword
and experience match analysis.

Identify the 10 most important keywords/skills in the job description.
Count how many appear in the resume (exact or close synonyms).
Base the ats_score primarily on that keyword hit rate, adjusted for
experience level match and job title alignment.

Do NOT default to a middle value — if the resume is a poor fit,
score it low (20–45). If it's an excellent fit, score it high (80–95).

Return only valid JSON:
{
  ats_score: integer 0–100,
  match_tier: 'Strong' | 'Partial' | 'Low',
  top_missing_keywords: [up to 3 crucial keywords absent from resume]
}

Resume Optimization Prompt (runs when you click Optimize):

You are an elite career coach and resume optimizer. Analyze the candidate's
resume against the job description and provide specific, actionable optimizations.

Return only valid JSON:
{
  match_score: 0–100,
  ats_breakdown: { skills_match, experience_match, title_match },
  top_3_changes: [exactly 3 specific changes],
  keywords_to_add: [5–8 crucial keywords],
  rewritten_headline: string under 15 words,
  rewritten_summary: 2–3 sentence professional summary
}

Both prompts run server-side through Replit’s AI integration — no Anthropic key required.

Quick-start checklist

1. Create a free Apify account and copy your API token

2. Open the Job Hunt Agent

3. Paste your Apify token

4. Enter your target job titles (comma-separated)

5. Set your preferred location (or leave blank)

6. Upload your current resume PDF

7. Hit Start Hunting and wait ~2 minutes for results

8. Sort by ATS score → focus on green matches

9. Click Optimize Resume on your top picks

10. Copy the rewritten sections into your resume and apply

Want to build this yourself? Here’s the Replit prompt.

I built this entire app on Replit using their AI agent. If you want your own copy — or want to customize it — just create a new Replit project and paste this prompt into the AI agent:

New to Replit? Use this link to get $10 in free credits.

Prompt:

Build a full-stack Job Hunt Agent app on Replit (React + Vite frontend, Express backend).

What it does: User uploads a PDF resume, enters an Apify API token and target job roles + location. The app scrapes LinkedIn for jobs posted in the last 24 hours, scores each posting against the resume using Claude AI (ATS keyword matching), displays a filterable/sortable results table, and lets users get a full resume optimization report for any job.

Frontend (React + Vite, dark SaaS design, deep navy/purple color scheme):

- Input form with: Apify token field (saved to localStorage), job titles field (comma-separated, each role searched independently), location field, PDF resume dropzone (parsed client-side with pdfjs-dist, text never leaves the browser)
- Loading state with progress text while scraping + scoring runs
- Results table with columns: Job Title, Company, Location, ATS Score (colored badge — green ≥72, yellow 42–71, red <42), Experience Level (parsed from description), Date Posted, Apply link
- Sort by any column, filter by experience level and minimum ATS score
- "Optimize Resume" button on each row that opens a side panel showing: skills/experience/title match breakdown, top 3 resume changes, keywords to add, rewritten headline, rewritten professional summary
- ATS scoring runs in batches of 5 jobs with a 1.5s delay between batches to avoid overwhelming the API
- A /thumbnail route that renders a fixed 1280×720 YouTube thumbnail (no nav/chrome) with: dark navy gradient background, subtle grid overlay, photo at public/images/abhi.jpeg on the right with glow effect, bold headline "I Built an AI Agent to Apply to Jobs For Me" with purple/blue gradient accents on key words, "Aspyre with Abhijay" channel branding, feature badges (Claude AI, ATS Scoring, LinkedIn Scraper, Resume Match), Episode 16 label

Backend (Express on port 8080):

- POST /api/search-jobs — takes { roles, location, apifyToken }. Splits roles by comma, searches each independently in parallel using Apify actor curious_coder~linkedin-jobs-scraper. For each role, passes 2 LinkedIn search URLs (start=0 and start=25, f_TPR=r86400 for 24h only) to get up to 50 results. Maps item.link as the job URL (not applyUrl which is external). Filters out jobs older than 24h by parsing relative dates ("2 hours ago", "1 day ago"). Deduplicates by URL. Falls back to a local data/cached-jobs.json if Apify returns 0 results.
- POST /api/score-ats — takes { resumeText, jobTitle, jobSnippet }. Uses Replit's Anthropic AI integration (no user API key needed) with Claude. Prompt instructs Claude to identify the 10 most important keywords in the JD, count matches in the resume, and return a genuine varied score (not defaulting to a middle value). Returns { ats_score, match_tier, top_missing_keywords }.
- POST /api/optimize-resume — takes { resumeText, jobDescription }. Returns { match_score, ats_breakdown: { skills_match, experience_match, title_match }, top_3_changes, keywords_to_add, rewritten_headline, rewritten_summary }.
- POST /api/fetch-jd — fetches full job description from a URL using Apify actor apify~website-content-crawler.
- All AI calls go through @workspace/integrations-anthropic-ai (Replit's managed Anthropic integration) — no Anthropic key field in the UI.

Misc:

- The Apify token is the only credential the user provides
- No salary filter should exclude jobs without salary info
- Include a data/cached-jobs.json fallback with ~5 sample LinkedIn jobs for when Apify returns nothing

That prompt will build the whole thing. You’ll have a working Job Hunt Agent in about 5 minutes.

Why this beats applying cold

Most people apply to jobs with the same generic resume and hope for the best.

This agent tells you before you apply whether your resume actually matches what the ATS is looking for.

The difference between a 45 and an 80 is usually 5–8 keywords — terms you probably know but didn’t think to include. The Optimize panel finds exactly those gaps.

It won’t get you the job. But it gets your resume past the filter so a human actually reads it.

And that’s the whole game.

If you are finding value in my AI action letters, please consider subscribing. It keeps me going.

If you know someone who’s job hunting right now — or about to start — please share this with them. This tool could save them hours every single week.

That’s it from me today.

Till next time. Stay tuned, as I will share the best resources from both my Harvard and Google networks to bring you the best. Let’s upskill together. Aspyre higher!

It’s March 2026. Hiring is picking up.

I still remember sitting in my apartment in December 2022, refreshing my inbox, waiting for a single callback. I had applied to over 100 jobs that month. I got zero interviews.

Not one.

And it wasn’t because I was unqualified. It was because I was doing it wrong. Mass applying. Generic resume. No strategy.

Fast forward to today — I’m a PM at YouTube/Google, I have a Harvard MBA, and I’ve helped thousands of people through this newsletter land real interviews.

The system I’m about to share with you is the same one that changed everything for me. I’ve distilled it into 5 actionable steps. No fluff. Just things you can do this week.

And here’s my promise: if you complete all 5 steps by end of April and still don’t land a single interview, I will personally get on a call with you.

That’s how confident I am.

Let’s get into it.

A quick note for my international readers: Some of you may have applied to the H-1B lottery. If you are stressed about it or this is your last chance, please explore H-1B alternatives, such as EB-1A or O-1 visas, especially if you have a STEM background.

I've partnered with Manifest Law to get you a free 15-minute consultation. These are legitimate paths that many people don't know about. If this applies to you, it's worth a conversation: Book your free consultation here.

Step 1: Fix Your Resume (Most Resumes Get Rejected in 7 Seconds)

Here’s the brutal truth — 75% of resumes are rejected by ATS before a human ever sees them.

Your resume isn’t being read by a person. It’s being parsed by software. If the format is wrong, it doesn’t matter how good your experience is.

The rules are simple:

• Single-column layout — multi-column layouts break ATS parsing

• No icons, tables, images, or graphics — ATS can’t read them

• Standard section headers — use exactly: Education, Experience, Skills, Projects

• Reverse chronological order — most recent first

• PDF format — always. You can use my template here. Subscribe to my Newsletter if you just downloaded it!

• One page — unless you have 10+ years of experience

Now fix your bullets. Every bullet point should follow this formula:

[Action Verb] + [What You Did] + [Quantified Result]

Here’s the difference:

• Weak: “Responsible for managing social media accounts”

• Strong: “Grew Instagram following from 2K to 50K in 6 months by implementing a data-driven content strategy, increasing engagement rate by 340%”

If your bullets don’t have numbers, rewrite them. At least 70% of your bullets should include a metric.

Step 2: Optimize Your LinkedIn Profile with This AI Prompt

Your LinkedIn profile is your digital first impression. Recruiters search LinkedIn before they look at your resume.

If your profile isn’t optimized, you’re invisible.

Here’s exactly what to do:

1. Go to your LinkedIn profile

2. Copy all the text — headline, about, experience, skills, everything

3. Open ChatGPT or Claude

4. Paste the prompt below along with your profile content

5. Apply the rewrites it gives you

The LinkedIn Optimization Prompt:

Here is the scraped content of a LinkedIn profile:
<profile>
{{PASTE YOUR LINKEDIN PROFILE CONTENT HERE}}
</profile>

The user's goal is: {{YOUR GOAL}}
(e.g., "attract brand sponsorships", "get recruited for senior PM roles",
"build thought leadership in AI")

Please analyze the profile across these dimensions and give specific
rewrites, not just suggestions:

1. Headline — Is it keyword-rich, role-specific, and value-forward?
   Rewrite it.

2. About/Summary — Does it tell a compelling story with a clear hook
   in line 1? Rewrite it.

3. Experience bullets — Are they outcome-driven with numbers? Flag
   weak ones and rewrite 2–3 examples.

4. Featured section — What should be pinned here given their goal?

5. Skills & Keywords — What's missing for SEO and
   recruiter/algorithm discoverability?

6. Creator/Posting signals — Any gaps in how they present their
   content or niche authority?

7. CTA — Does the profile have a clear next step for visitors?

Format your output as:
[Section] → [Issue] → [Rewrite or Recommendation]

Step 3: Stop Finding Jobs Manually — Automate Job Discovery with AI

Here’s the truth about job boards: by the time you see a posting on LinkedIn or Indeed, hundreds of people have already applied. The early applicants get interviews. Everyone else gets ghosted.

The solution? Automate job discovery so you see postings within 24 hours of being listed.

I built an automation using Apify and n8n (~$7/month, self-hosted) that:

• Scrapes LinkedIn for jobs posted in the last 24 hours

• Filters by role, location, and industry

• Returns the top 5 most recent listings per search

• Delivers results to you daily — no manual searching

But here’s the real move — don’t just apply through the portal.

Find the actual hiring managers on LinkedIn using these search patterns:

• Pattern A (Direct Intent): Search for “I’m hiring”, “looking for a”, “open role on my team”

• Pattern B (Call to Action): Search for “DM me”, “send me your resume”, “drop your portfolio”

• Pattern C (Team Growth): Search for “growing the team”, “excited to announce”, “just opened a req”

A direct message to a hiring manager is 10x more effective than an Easy Apply.

I wrote detailed setup guides for both of these:

• AI Agent That Job Hunts While You Sleep — Full n8n + Apify pipeline tutorial

• Use LLMs in Your Job Search — LinkedIn search patterns and prompts for finding hiring managers

Step 4: Auto-Tailor Your Resume for Every Job (One Click)

Here’s the problem. You found the perfect job listing. But your resume says “project management” and the job description says “program management.”

ATS doesn’t know those are the same thing. Your resume gets filtered out.

The fix is to tailor your resume to each job description. But doing that manually for every application? Brutal.

So I automated it.

How the workflow works:

1. Job descriptions are fetched automatically from your Step 3 pipeline

2. AI extracts keywords — skills, tools, qualifications, action verbs

3. Your base resume is rewritten to match the job’s language — without inventing experience you don’t have

4. A new Google Doc is created for each tailored resume

5. A summary email is sent to you with links to every tailored resume and the job posting

You wake up to an email with your tailored resumes ready to submit.

• n8n — Workflow orchestration (~$7/month). Use code “ABHIJAY” to get additional 10% off. Link here.

• Apify — LinkedIn job scraping (Free tier)

• Google Gemini API — Resume rewriting (Free tier)

• Google Docs — Tailored resume docs (Free)

• Gmail — Summary emails (Free)

Full setup walkthrough: AI Agent That Job Hunts While You Sleep

Step 5: Post on LinkedIn — It’s Your Secret Weapon

If you don’t market yourself, who will?

Posting on LinkedIn during your job search does 3 things:

1. Signals to recruiters that you’re active and engaged in your field

2. Creates inbound opportunities — hiring managers find you

3. Builds your professional brand so when someone Googles your name, they find substance

You don’t need to go viral. You just need to start.

Post 3x per week. Here are some ideas to get you going:

The Reintroduction:

• “I’ve been lurking on LinkedIn for 3 years. Here’s what I’ve learned just by reading.”

• “I’ve never posted here before. So let me finally introduce myself properly.”

Lessons Learned:

• “The one career lesson I had to learn the hard way.”

• “3 things I know now that I wish someone told me on day one of my career.”

Hot Takes:

• “Cold emails work better than job boards. Change my mind.”

• “The soft skills that companies say they want vs. the ones they actually reward.”

Industry Insights:

• “The part of the AI conversation that I think is missing from most discussions.”

• “3 skills that will matter most in [your field] over the next 5 years.”

Tools & Productivity:

• “How I use AI in my actual workflow — not the hype version, the real one.”

• “The one tool that made my work dramatically more efficient this year.”

I’ve put together 90 post ideas across 9 categories. Reply to this email and I’ll send you the full list.

90 LinkedIn Post Ideas for Your First (Real) Post

Here’s a full bank organized by format and intent — pick whatever fits your voice.

🔄 The “I’m Back” / Reintroduction (1–10)

1. “I’ve been lurking on LinkedIn for 3 years. Here’s what I’ve learned just by reading.”

2. “I’ve never posted here before. So let me finally introduce myself properly.”

3. “I’ve consumed content on LinkedIn for years. Time to contribute something.”

4. “Most of my career wins happened while I was silent on here. Here’s a quick recap.”

5. “I’m not a natural poster. But I have things worth saying. Starting today.”

6. “If you’ve wondered who the quiet person in your network is — that’s been me. Until now.”

7. “I used to think posting on LinkedIn was for people with big titles. I was wrong.”

8. “I kept drafting posts and deleting them. This one’s staying up.”

9. “I’ve been at [company] for X years. Here’s what I wish I’d shared sooner.”

10. “I told myself I’d post ‘when the time was right.’ Turns out there’s no perfect moment.”

💡 Lessons Learned (11–20)

11. “The one career lesson I had to learn the hard way.”

12. “3 things I know now that I wish someone told me on day one of my career.”

13. “The best piece of advice I ever got — and why I ignored it for years.”

14. “What I learned in my first 90 days at my current job.”

15. “A mistake I made early in my career that actually shaped everything.”

16. “The thing nobody tells you about [your industry].”

17. “I spent 5 years trying to be the smartest person in the room. Here’s what changed.”

18. “What I learned from a boss I absolutely hated.”

19. “The meeting that changed how I think about my career.”

20. “One habit that made me significantly better at my job.”

🎯 Hot Takes & Contrarian Views (21–30)

21. “Unpopular opinion: hustle culture is making us worse at our jobs.”

22. “We over-celebrate busyness. Here’s what I think actually matters.”

23. “Your resume is not your story. Here’s what I think is.”

24. “Cold emails work better than job boards. Change my mind.”

25. “The ‘follow your passion’ advice is incomplete. Here’s the missing part.”

26. “Work-life balance is the wrong framing. Here’s a better one.”

27. “Networking events are mostly a waste of time. Here’s what isn’t.”

28. “I don’t think the 10,000-hour rule applies the way we think it does.”

29. “The soft skills that companies say they want vs. the ones they actually reward.”

30. “Being likable at work matters more than being competent. (Let’s talk about it.)”

📖 Personal Story Arc (31–40)

31. “I almost quit my career in [field] 3 years ago. Here’s what stopped me.”

32. “My career didn’t go according to plan. Here’s the version that actually happened.”

33. “I got laid off. Here’s what that week looked like — and what came next.”

34. “I changed industries at [age]. Everyone thought I was crazy. Here’s how it went.”

35. “The promotion I didn’t get — and why I’m grateful for it now.”

36. “I relocated for a job. It didn’t work out. Here’s what I learned.”

37. “From [starting point] to [current role] — the version nobody puts on their resume.”

38. “I said yes to something terrifying at work. Here’s what happened.”

39. “The conversation that completely changed my career trajectory.”

40. “How a random connection on LinkedIn led to [a major opportunity].”

📊 Industry Insight & Trends (41–50)

41. “Something I’m seeing shift in [your industry] that nobody’s talking about enough.”

42. “3 skills that will matter most in [your field] over the next 5 years.”

43. “Why [a common industry practice] is overrated.”

44. “The part of the AI conversation that I think is missing from most discussions.”

45. “What companies in [industry] keep getting wrong about [topic].”

46. “Here’s how [your field] has changed in the last 3 years — and what that means for professionals.”

47. “The role that didn’t exist 5 years ago that every company now needs.”

48. “An outdated belief in [your industry] that’s still being taught like it’s gospel.”

49. “Why I think [emerging trend] is more important than most people realize.”

50. “The thing that separates good companies from great ones in [your industry].”

🤝 People & Management (51–60)

51. “The best manager I ever had did one thing differently.”

52. “What I look for when I’m hiring — and it’s probably not what you’d expect.”

53. “I’ve interviewed hundreds of people. The candidates who stand out all do this.”

54. “The type of teammate that makes every project better.”

55. “Something I noticed great leaders do that average leaders don’t.”

56. “Why I started asking ‘what do you need from me?’ more than ‘how’s it going?’”

57. “The feedback conversation that was hard to give — and why I’m glad I did.”

58. “How I think about building trust with a new team quickly.”

59. “The most underrated quality in a professional: [your answer].”

60. “What managing people taught me about myself.”

🛠️ Tools, Systems & Productivity (61–70)

61. “The one tool that made my work dramatically more efficient this year.”

62. “How I structure my week to protect time for deep work.”

63. “My personal framework for making decisions at work.”

64. “I started doing [small habit] 6 months ago. Here’s what changed.”

65. “The way I take notes has completely changed how I retain information.”

66. “How I use AI in my actual workflow — not the hype version, the real one.”

67. “The system I use to track my wins throughout the year (for performance reviews).”

68. “What I do in the first 30 minutes of every workday.”

69. “The question I ask at the end of every project to get better over time.”

70. “Why I started writing a ‘failure log’ — and what it’s taught me.”

🌍 Values, Purpose & Identity at Work (71–80)

71. “The moment I realized my job title isn’t my identity.”

72. “What I believe about work that I think most companies get wrong.”

73. “Why I turned down a higher-paying job — and what that decision taught me.”

74. “The part of your career that has nothing to do with your resume.”

75. “Something I protect no matter how busy work gets.”

76. “What ‘meaningful work’ actually means to me.”

77. “Why I stopped optimizing for status and started optimizing for [something else].”

78. “The values I want to still have in 20 years — and how I try to practice them now.”

79. “A thing I’m genuinely proud of at work that wasn’t a promotion or raise.”

80. “What I think about when I’m deciding whether an opportunity is right for me.”

🎓 Learning & Growth (81–90)

81. “The book that most changed how I think about work. (And the idea that stuck.)”

82. “What I learned from 30 days of [challenge or experiment].”

83. “I took a course on [topic] expecting one thing and got something completely different.”

84. “The mentor who shaped my career — and the one piece of advice they kept repeating.”

85. “What I’m actively trying to get better at this year.”

86. “The question I asked that unlocked a completely new way of thinking for me.”

87. “Something I believed at 22 about work that I’ve completely unlearned.”

88. “The podcast/newsletter/resource that I’d recommend to everyone in [field].”

89. “What I’m learning right now — and why it feels uncomfortable.”

90. “The thing I realized I was bad at — and what I did about it.”

Pro tip: Posts that perform best on LinkedIn tend to start with a bold first line, use short punchy paragraphs (1–2 sentences), and end with a question or call to action. Pick one from a category that feels most authentic to you right now — you don’t need to be loud, just specific.

Your Quick-Start Checklist

• Step 1: Create your ATS-friendly resume using the formatting rules above

• Step 2: Run your LinkedIn profile through the AI optimization prompt

• Step 3: Set up Apify + n8n for daily job scraping (setup guide)

• Step 4: Configure the ATS resume tailoring automation (setup guide)

• Step 5: Pick 3 post ideas and publish your first LinkedIn post this week

Remember my promise. Complete all 5 steps by end of April. Show me proof. If you still don’t have an interview, I’ll get on a call with you.

If you know someone who is job hunting right now, please share this with them. I promise these 5 steps will make a real difference.

That’s it from me today.

Till next time. Stay tuned as I will share best resources from both my Harvard and Google networks to bring the best to you. Let’s up skill together. Aspyre higher!