The artifacts, evals, and decision frames AI PMs use to ship features without breaking trust.
The preflight that kills bad ideas before they consume engineering time.
Whether you're sketching an idea or operating in production, the artifacts feed the same decision: ship, hold, or stop.
If you only steal three things from this playbook, take these. Opportunity → PRD → Launch gate. Everything else hangs off them.
Every serious AI product moves through this loop. Each step has a single decision it unlocks. Skip one and the next one breaks.
Is AI actually the right tool here?
Spec the AI job and what 'good' means.
Go / no-go with evidence, not vibes.
Catch drift, cost, and silent regressions weekly.
Observability feeds the next opportunity brief. The loop is the product. Run it every cycle.
Read the loop guideEach one includes an opportunity brief, PRD, eval plan, launch gate assessment, and a scored readiness recommendation.
“AI drafts replies for Tier 1 agents. Strong job definition, but evals, observability, and a low-confidence fallback have to land before pilot.”
“Auto-summarize calls and update CRM fields. The risk is silent over-writes, so we need accept/edit telemetry and a rollback path before scale.”
“Patient intake is the highest-blast-radius workflow in the product. Evidence does not support a pilot. The recommendation is to stop and rescope.”
Most of these are PM jobs now. Each one has a template you can steal.
Each one is short, opinionated, and uses real examples. Pick the phase you're in and start there.
Building evals that catch real failures, including the ones you miss in demos.
How to spec agents vs. chatbots vs. copilots. Autonomy levels and tool boundaries.
Treating prompts as product surfaces. Versioning, regression, and ownership.
Tool boundaries, autonomy, rollback, trajectory evals, cost ceilings, handoff.
It is not a failure state. It is what you say when the blast radius is larger than the team's ability to measure, review, roll back, or operate the AI safely.
MIT licensed. Built in the open. If it makes your next AI review less hand-wavy, a GitHub star tells me it's worth maintaining.