Technical product management for the AI PM

Product sense tells you what to build. Technical product sense tells you what the system will let you build. Technical product management is the discipline that turns those judgments into shipped software. It covers the role, the artifacts, the rituals, and the release machinery that carry an idea from a hunch to a feature running reliably in production. It's the operating system of the PM job — the part you're actually evaluated on when the quarter ends.

For APMs and PMs moving into AI product management, the craft matters double. AI features are harder to spec, because behaviour is probabilistic. They're harder to estimate, because quality is discovered, not designed. They're harder to launch, because a model can regress silently. And they're harder to measure, because the interesting failures don't throw errors. Every lesson here teaches the general practice first, then shows exactly what changes when there's a model in the build. Each lesson ships a diagram you can redraw on a whiteboard.

The knowledge graph

The craft is a loop that runs every quarter — with the role at the center and the AI capstone bending every station:

• The technical PM role — what a technical PM actually owns, the PM ↔ TPM ↔ EM spectrum, and where your leverage comes from.
• Discovery to delivery — the product development lifecycle: dual-track discovery and delivery, and the loops that keep them honest.
• Specs, PRDs & RFCs — the document stack: writing a PRD engineers respect, reading an RFC, and requirements that survive contact with reality.
• Prioritization & roadmaps — RICE, cost of delay, and the Kano model; roadmaps as bets, not promises; and how to say no.
• Working with engineering — sprints, rituals, estimates, and the trust economy between PMs and engineers.
• Metrics & experimentation — the metric tree, instrumentation as a requirement, and A/B testing without fooling yourself.
• Launches, rollouts & migrations — feature flags, progressive delivery, rollback plans, and retiring the old thing safely.
• Incidents & postmortems — severity, incident command, honest communication, blameless learning — and the AI extension (quality incidents, kill switches).
• Technical product management for AI — the capstone: eval-driven development, probabilistic acceptance criteria, model upgrades as migrations, and the data flywheel.

Each lesson pairs the craft with a 🎯 For the AI PM briefing — how the practice bends when the product has a model in it — and a diagram to make it concrete.

Connects to other tracks

• Product sense — the judgment that decides what's worth shipping.
• Technical product sense — reading the system you're shipping on.
• Evals: golden sets, adversarial, LLM-as-judge — the measurement layer under eval-driven development.
• Versioning & migration (Flowable) — model upgrades as migrations, made concrete.

📌 Close out the module: Recap & real-world examples.