The complete, verified, no-hallucination guide to becoming an AI Product Manager, and passing the interviews.
From "AI is magic" → to shipping eval-gated AI products → to offers at frontier labs.
Foundations · Technical depth · Product craft · Building with AI · 7 domains · Sourced case studies · 150+ interview questions · Career
🗓️ All facts verified July 2026 · every stat sourced · every myth flagged · see our verification standard ↓
New to this repo? → 00-start-here picks your learning path in 60 seconds.
| If you are... | Your path | Time |
|---|---|---|
| 🌱 Aspiring PM / student | Path A: Full Foundation | 12-16 weeks |
| 🔄 PM moving into AI | Path B: AI Upskill | 6-8 weeks |
| 🧑💻 Engineer / Data Scientist → PM | Path C: Product Craft | 6-8 weeks |
| 👔 Senior PM / Director leading AI | Path D: AI Leadership | 4-6 weeks |
| 🎯 Interviewing in 30 days | Path E: Interview Sprint | 2-4 weeks |
mindmap
root((AI PM))
00 Start here
5 learning paths
Skill self-assessment
01 Foundations
What AI PMs actually do
ML fundamentals
LLM fundamentals
100+ term glossary
02 Technical depth
Prompt & context engineering
RAG
Fine-tuning
Agents
Evals ⭐
Model selection & cost
03 Product craft
AI product strategy
Discovery for AI
The AI PRD
UX patterns & trust
Metrics
Responsible AI
04 Build with AI
Verified tool stack
5 portfolio projects
05 Domains
Healthcare · Fintech
E-commerce · Enterprise
Consumer · DevTools · EdTech
06 Case studies
Copilot · Netflix · Duolingo
Klarna · 10 sourced failures
07 Interview prep
Loops by company
Answer frameworks
150+ question bank
Worked mocks
08 Career
Who's hiring
Compensation
Transition guides
Resume & portfolio
09 Resources
Courses · Books · Tools
Newsletters · Regulation · Stats
Five learning paths, the AI PM skill map, and the three rules of this repo.
| Page | What you get |
|---|---|
| What AI PMs actually do | The 4 archetypes, AI PM vs. classic PM, a week in the life, skills self-assessment, myths debunked |
| ML fundamentals | Programming-with-examples, precision/recall as a product decision, the "should this be ML?" decision tree |
| LLM fundamentals | Next-token prediction → every consequence that matters; the 3-stage training pipeline; explain-it-to-an-exec cheat sheet |
| Glossary | 100+ terms grouped by theme, with the ~30 you must speak fluently marked |
02 · Technical Depth (the interview differentiator)
| Page | The question it answers |
|---|---|
| Prompt & context engineering | Why the prompt is product surface, and how to version + eval it |
| RAG | How to ground AI in your data, and the retrieval-vs-generation debugging instinct |
| Fine-tuning | When to climb the adaptation ladder, and why the answer is usually "not yet" |
| AI agents | The autonomy dial, compounding errors, why agent safety is permissions not prompts |
| Evals ⭐ | Evals are the new PRD: the single most differentiating AI PM skill |
| Model selection & cost | The unit-economics math that gets you hired; verified July 2026 price map |
| Page | What it gives you |
|---|---|
| AI product strategy | Moats, the wrapper question, the capability curve, business-model patterns with real anchors |
| Discovery for AI | The AI-fit screen, the data audit, the feasibility spike |
| The AI PRD | Copy-paste template + a fully worked example |
| UX patterns for AI | Trust engineering: citations, confidence, failure UX, friction-as-a-feature |
| Metrics | The 3-layer stack, AI-specific traps, north star + guardrails |
| Responsible AI | Bias, privacy, prompt injection, governance: as product work, not legal review |
The verified 2026 tool stack + 5 portfolio projects that answer "what have you built?", from a version-scored prompt to an agent you red-teamed yourself.
Same craft, different physics. Healthcare · Fintech · E-commerce & Retail · Enterprise SaaS · Consumer & Social · Developer Tools · EdTech
GitHub Copilot · Netflix Recommendations · Duolingo · Klarna's AI Assistant · The Failures Hall of Lessons: 10 sourced disasters, one transferable lesson each.
| Page | What's inside |
|---|---|
| Loops by company | OpenAI, Anthropic, Google, Meta (incl. the new AI round), Amazon, Microsoft, Perplexity |
| Answer frameworks | AI-adapted structures for design, technical, metrics, strategy, behavioral + the live-prototyping round |
| Question bank | Design · Technical · Strategy · Execution · Behavioral. Sourced 🔵 vs. practice ⚪ |
| Mock scenarios | Three full interviews worked out loud, with commentary on why each move scores |
Who's hiring (verified postings, 4 tiers, incl. India) · Compensation (levels.fyi + posted ranges, US & India) · Transition guides (4 paths in) · Resume & portfolio
Courses · Books · Tools · Newsletters, podcasts & communities · Regulation & governance · Stats & benchmarks
"Research 1000s of places and make it the best. No hallucinations." (the brief this repo was built to)
Most AI PM content on the internet is confidently wrong: invented statistics, dead tools, prices from two pricing models ago, interview questions someone made up. This repo took the opposite approach.
| What we did | Why it matters |
|---|---|
| Every stat traced to a primary source: papers, SEC filings, tribunal records, company announcements, official pricing pages | You can cite these in an interview without getting caught |
| Every fact date-stamped | AI facts have a half-life; you know exactly how fresh yours are |
| Unverifiable claims omitted or flagged |
We deleted prices we couldn't confirm rather than guessing |
| A dedicated myths list of popular numbers that did not survive checking | So you don't repeat the internet's favorite fabrications |
| Interview questions labeled by source: 🔵 reported for a named company vs. ⚪ practice | No pretending invented questions are real ones |
Examples of what that discipline caught: the "Klarna fired 700 people and replaced them with AI" story is false as usually told · the Netflix "$1B / 80%" stats are self-reported and a decade old · Humanloop shut down in 2025 (still recommended by plenty of listicles) · "Copilot writes 46% of code" has no stable primary source. Details in stats-and-benchmarks.md.
Volatile by nature, re-verify before citing: model names & API prices, tool pricing, M&A status, regulation timelines. The tools directory flags exactly which facts rot fastest.
Five opinions, defended throughout:
- Evals are the new PRD. If you internalize one idea, make it this one. Most people calling themselves AI PMs have never written an eval; that gap is your opportunity.
- Build before you badge. One shipped prototype beats three certificates. Five projects, a weekend each.
- AI PM is a superset of PM, not a different species. Discovery, prioritization, and storytelling still decide everything (plus four new muscles).
- The model is the engine, not the car. Moats live in workflow, data flywheels, distribution, and trust (strategy).
- Failure literacy is a hiring signal. The failures section exists because "what could go wrong?" is where senior candidates separate.
|
Explain it in one line
|
The five interview reflexes
|
Found a broken link, a stale price, a better resource, or a fact we got wrong? CONTRIBUTING.md explains the one rule: claims need sources. A PR that adds a great resource with a link and a date gets merged; a PR that adds an unsourced statistic does not.
Especially wanted: newer sourced interview questions (with attribution), domain pages we haven't written (gaming, logistics, legal, climate, agritech), and corrections to anything that aged badly.
CC BY 4.0. Use it, remix it, teach with it. Attribution appreciated.
⭐ Star this repo if it helped you. It's how other PMs find it.
Built for everyone who wants to build AI products that actually work.