Use this when you need to know which artifact comes next. The playbook is a decision system: each artifact should unlock a specific product decision or make a blocker visible.
| Phase | Artifact | Owner | Decision unlocked | Next step |
|---|---|---|---|---|
| Idea screening | Before you vibe code | PM | Is the idea clear enough to investigate? | If unclear, write an opportunity brief. If clear, define the AI job. |
| Opportunity assessment | AI opportunity brief | PM with cross-functional partners | Is AI worth pursuing, and is the idea coherent across user, AI job, data, evals, risk, cost, and human control? | If yes, write the PRD. If no, use a non-AI path or stop. |
| Product definition | AI PRD | PM with design and engineering | What does the AI do, for whom, under what constraints and risks? | Create the eval plan, review workflow, cost model, observability plan, and launch gate. |
| Engineering handoff | Optional AI build brief | PM with engineering lead | Is the work scoped enough to build or prototype? | Build the smallest useful version or spike. |
| Quality definition | AI eval plan | PM with domain expert and ML/engineering | How will the team know whether output is good? | Run baseline evals and update the launch gate. |
| Failure discovery | Error analysis | PM or domain expert | What is actually going wrong in traces? | Add recurring failures to evals and risk register. |
| Risk ownership | AI PRD risk table | PM with legal, security, design, engineering | What could go wrong, how bad is it, and who owns it? | Mitigate high-severity risks before launch gates. |
| Human control | Human review workflow | PM with design and operations | Where must humans approve, edit, reject, or undo AI output? | Design review UI and audit trail. |
| Unit economics | AI cost model | PM with engineering and finance | Is the workflow economically viable at expected usage? | Set cost ceilings and observability targets. |
| Prompt management | Optional prompt change record | PM with engineering or AI lead | Is a prompt change justified, tested, reviewed, and rollbackable? | Roll out through staging, shadow mode, canary, or production. |
| Meeting review | Optional AI PM review checklist | PM | Are roadmap, design, engineering, legal, or launch concerns blocking progress? | Proceed, hold, or recommend do not launch. |
| Launch decision | Launch gate checklist | PM with accountable approvers | Can the product enter pilot, limited production, or scale? | Document the decision and conditions. |
| Production monitoring | AI observability plan | PM with engineering/data | How will the team detect drift, cost spikes, failures, and user trust loss? | Review signals on cadence and feed failures back into evals. |
| Operating loop | Weekly post-launch review | PM with product team | Did the product perform as expected after launch? | Continue, roll back, hold expansion, or update the roadmap. |
Two common routes
From idea to pilot: preflight -> opportunity brief -> PRD -> eval plan -> human review workflow -> cost model -> observability plan -> launch gate.
From prototype to confidence: error analysis -> eval plan -> PRD risk table -> observability plan -> optional AI PM review checklist -> launch gate.
Rule of thumb
If an artifact does not help the team make a better decision, skip it. If a decision feels subjective or political, add the missing artifact that would make the evidence visible.