to-prd
Installation
SKILL.md
To PRD
Turn .ai-workflow/work/<feature-slug>/decision-record.md, status ready-for-prd, into .ai-workflow/work/<feature-slug>/prd.md. You draft it yourself. The PRD stays at the decision level: no fresh architecture choices, files, code, patterns, or test mechanics.
- Reuse the record's decision IDs and source index, and respect already decided choices. Recheck a source only when its locator or revision is missing, or a fact changed or is disputed. A missing material decision or a conflict goes back to
interview-mewith the exact question and the conflicting sources; never write an approvable PRD from a guess. - Write the draft with
# PRD: <feature name>and**Status:** draft, and these sections:- Problem and goals.
- User stories or journeys.
- Observable acceptance outcomes, including important failure behavior.
- Settled product, domain, integration, and data decisions.
- Security decisions, or
No exposed security surface. - Rollout and rollback constraints.
- Out of scope, assumptions, and remaining risks.
- Decision IDs for each outcome and constraint, linked to the record rather than copied.
- Use the glossary's terms and define any new one once. Where a rule stays ambiguous, add a short scenario naming the actor, record, and outcome.
- Present the path and the material choices. Revise from user corrections. Mark it
approvedonly on explicit acceptance, recording the user's actual words or the host's message locator; silence does not approve it. An--autohandoff frominterview-mestill needs this approval.
The approved PRD is the only input to-tasks accepts.