Ship a real AI feature. Then answer for it.

Have your hosted workspace live and your repo ready before session 1 so no live hour is spent on installation.

• Your hosted AI workspace goes live
• GitHub account and portfolio repo scaffold from template
• Confirm your case: B2B Procurement and Supply Chain Operations
• Post an intro in the cohort channel

You finish with: Hosted workspace live, portfolio repo scaffolded for the case

Tools: GitHub, your hosted workspace, Vercel or Netlify, Figma (free tiers)

Learn to see an AI product as a set of design decisions and failure modes rather than a demo.

• Anatomy of an AI feature: model, context, interface, fallback
• Live teardown of two shipped products in your vertical
• Build the failure catalog: what breaks, how visibly, who pays
• Agent-readiness lens: is this product legible to an agent, not just a human
• Peer swap and critique

You finish with: Teardown doc + failure catalog + agent-readiness note

Tools: Chosen products, teardown template

Compress weeks of qualitative discovery into hours, and leave with a workflow you can rerun on any future problem.

• Sourcing evidence: reviews, tickets, forums, transcripts
• AI-assisted synthesis without hallucinated themes
• Bias controls and disconfirming evidence
• Save the workflow as a reusable skill / slash command / prompt library
• Select and defend one problem

You finish with: Insight report + chosen problem + reusable discovery skill

Tools: Claude Code or equivalent, transcript and review corpus

Answer 'can this actually be built, with what, on what data' before writing a single requirement.

• Prompt vs RAG vs fine-tune decision tree
• The latency, quality and cost frontier
• Bake-off: run one task across three models and score it
• Data readiness: does it exist, who owns it, is it labeled, may we use it
• Feed both outputs into tomorrow's scorecard

You finish with: Model shortlist + data readiness check

Tools: 2-3 frontier and small models, eval prompt set

Turn a validated problem plus a feasibility read into a decision-ready brief a skeptical exec would sign.

• The 2-page brief structure
• Feasibility scorecard, now grounded in session 3 evidence
• Risk register: what could go wrong, likelihood, mitigation, owner
• Regulatory literacy: EU AI Act risk tiers and NIST AI RMF in 30 minutes
• Which risks are yours vs the model provider's

You finish with: 2-page brief + feasibility scorecard + risk register

Tools: Brief template, scorecard, risk register template

Treat the system prompt, retrieved context and tool descriptions as the product specification, and version them like code.

• The system prompt as spec: structure, constraints, refusals
• Context engineering: what goes in the window, what does not, and why
• Retrieval design: chunking, grounding, citation, freshness
• Context budget and the cost of every token
• Break your own spec and log every failure

You finish with: Versioned context spec + retrieval design + break log

Tools: Version-controlled prompt repo, small document corpus

Move from a single-turn prompt to an agent that plans, calls tools and hands off work under human control.

• The agent loop: plan, act, observe, repeat
• Tool and function calling: schemas, errors, retries
• MCP: connecting an agent to real systems
• Memory: what persists, what should not
• Orchestration: sequential vs parallel, handoffs between agents
• Approval gates and where a human must sit in the loop

You finish with: Working agent calling 2+ tools, with a handoff and an approval gate

Tools: Claude Code, MCP server, 2 external tools or APIs

Ship something a stranger can open, driven by the context spec and agent built in the previous two sessions.

• From spec to scaffold
• Wiring the agent into a real interface
• State, auth and the minimum viable backend
• Deploy to a public URL
• Repo hygiene: README, env example, license

You finish with: Live prototype URL + public repo

Tools: Claude Code or Cursor, Vercel or Netlify, Supabase free tier

Design what happens when the model is wrong, slow, uncertain or misused, and prove you tried to break it.

• The trust pattern library: confidence signalling, citations, undo, graceful degradation, AI labeling
• Approval flows and progressive autonomy
• Designing three sad paths: wrong answer, no answer, harmful answer
• Applied red team: adversarial prompts against your own prototype
• Bias spot-check on your own outputs

You finish with: Clickable flows incl. 3 sad paths + guardrail pattern set + red-team log

Tools: Figma, prototype from session 7, red-team prompt set

Replace 'it feels better' with a repeatable measurement your stakeholders can audit.

• From failure catalog to failure taxonomy
• Building an eval set that reflects real usage
• Rubrics, LLM-as-judge and where it misleads
• Human review: sampling, inter-rater agreement
• Run baseline, change one thing, run again, report the delta

You finish with: Eval set + failure taxonomy + before/after report

Tools: Eval harness, labeled example set

Keep the thing working after launch, when the inputs stop looking like your test set.

• Offline vs online evaluation
• Tracing and logging: what to capture on every call
• Drift: how it shows up and how you notice
• Regression gates in CI
• Cost and latency as monitored metrics
• What an AI incident looks like and who is on call

You finish with: Tracing enabled + online eval dashboard + drift alert

Tools: Tracing/observability tool free tier

Connect model behavior to business outcomes, and know what a single user costs you.

• The metrics tree: model quality to product outcome to business result
• Unit economics: cost per task, per user, per resolved case
• Levers: caching, routing, small models, context trimming
• Total cost of ownership beyond API spend
• Build vs buy vs partner
• Experiment design on a stochastic system: valid A/B, guardrail metrics, sample size intuition

You finish with: Metrics tree + cost model + experiment plan

Tools: Cost model template, usage logs from sessions 7-10

Write the document that gets the thing funded, and survive the questions that follow.

• The AI PRD: what differs from a standard PRD
• Rollout: beta gating, staged release, adoption metrics, kill criteria
• The 5-slide business case
• Live pitch and challenge: present, then take hostile questions from peers playing engineering, finance and legal
• Rewrite based on what broke

You finish with: AI PRD + 5-slide business case + rollout plan

Tools: PRD template, slide template

Package the work from every session into one credible piece of evidence, and be able to talk about it under pressure.

• Assembling the artifact chain into a portfolio site
• The 3-minute demo video
• The interview narrative: problem, decision, evidence, trade-off
• Five AI PM questions your portfolio already answers
• Demo day: present to the cohort and invited reviewers
• Capstone rubric review and certification

You finish with: Live portfolio site + demo video + interview narrative

Tools: Portfolio template, screen recorder

Yes, and you're the reason this workshop exists. AI PM is young enough that nobody has ten years in it. Interviewers look for two things: have you shipped something real with AI, and can you reason about evals, tradeoffs, and product decisions under pressure. The workshop manufactures both.

Yes. Everything hands on happens in plain English with Claude, inside your own hosted workspace. The skill being tested is product judgment: what to build, how to know if it works, what to do when the model is wrong. You write no code at any point.

A shipped AI artifact you built yourself in B2B Procurement and Supply Chain Operations, an eval set you designed (the most asked about item in AI PM interviews), a one page product memo in the format hiring managers read, and answers pressure tested against our free bank of 1,000 AI PM interview questions. Artifacts, not notes.

Roughly 11 hours a week across 7 weeks: two live sessions plus published async work, so nothing is a surprise homework load. Every session is recorded and shared within 24 hours, along with the templates used in it, so a missed session never costs you a week.

No hidden costs. No extra tools to buy. Your hosted workspace includes all the model usage for the program, so there's no separate API bill and no ChatGPT or Claude subscription to buy.

India: UPI, cards, and netbanking. Global: cards.

Full refund within 7 days, no forms to fight and no questions that matter. You risk a week of attention, not your money.

Yes. The portfolio artifact, memo format, and interview drills transfer to any PM interview. AI PM is simply where the gap between demand and credible candidates is widest, which makes it the smartest door to walk through first.