Use this to turn PM intent into engineering-ready work. Compatible with the GRIT framework for AI-assisted engineering delivery.
Upstream: pull the problem, AI job, and constraints from the AI PRD. Parallel: reference the eval plan for test requirements and the launch gate checklist for ship criteria.
Premise check
Problem: One sentence. What user problem does this solve?
Smallest useful version: What is the narrowest scope that still delivers value?
Does this already exist? Have you checked if an off-the-shelf model, API, or existing feature already handles this?
Null case: What happens if the AI produces nothing? What does the user see? Is that acceptable?
Behavioral contract
Must do:
Must not do:
When uncertain:
Inputs and outputs
Inputs:
| Field | Type | Source | Required | Validation |
|---|---|---|---|---|
Outputs:
| Field | Type | Always present | Fallback |
|---|---|---|---|
Edge cases
| Case | Example input | Expected behavior |
|---|---|---|
Non-goals
Dependencies and prerequisites
| Dependency | Status | Owner | Blocker if missing? |
|---|---|---|---|
Performance requirements
- Latency target: e.g., p50 < 2s, p95 < 5s, hard timeout at 8s
- Throughput: e.g., must handle X concurrent requests
- Cost ceiling: e.g., < $0.05 per task
Eval and test requirements
- Golden set accuracy: target
- Regression test: what specific behavior must not degrade?
- Adversarial test: what inputs should the AI handle gracefully?
- Safety test: what outputs are unacceptable regardless of other scores?
Risk and hardening
| Risk | Mitigation | Detection |
|---|---|---|