The 7-step playbook to build AI products that ship.

Pain × Frequency × Feasibility

Before you touch a prompt or a model, score the problem on three axes. If you can't score all three high, you have a science project — not an AI opportunity.

The Assumption Ladder

Every AI idea rests on assumptions about the user, the data, the model, and the business. Most PMs test only the top one and skip the ones underneath that actually kill projects.

Scope the failure, not just the feature

A traditional PRD defines what the feature does. An AI PRD also defines what happens when it's wrong — because it will be. No 'acceptable failure modes' section means it's not ready for engineering.

Design for trust before delight

Every AI UX decision should answer one question: does this help the user calibrate how much to trust the output? Confidence, reversibility, and explainability are the actual design problem.

Ship the thinnest reliable slice

Don't build the full agentic workflow before shipping anything. Find the thinnest version reliable enough to trust, ship it, and expand only after usage proves the foundation holds.

Eval before launch, not instead of launch

An eval set isn't a QA checkbox — it's what tells you whether you're allowed to trust your own feature. No labeled eval set with adversarial cases and a human baseline means you're guessing, not shipping.

The 4-layer metrics stack

Offline eval quality feeds in-product AI quality, which feeds user behavior, which feeds business impact. Most teams only report the business number and never prove the two are connected — that gap is why AI investment gets cut.