garyld1962 Execute PRD
Read the sibling validate-plan skill's requirements workflow and plan contract.
- Resolve one source and the repository command contract; audit current code and tests.
- Apply the automated readiness score before planning. A requirements interview is offered for gaps, never silently launched in an unattended run.
- Preserve requirement IDs, invariants and closed decisions. Use design-twice for material unresolved architecture in feature/refactor work.
- Author tasks with depends_on, write_scope, milestone_end and mechanical acceptance. Preserve all requirements while optimizing dependencies and shared ownership.
- Validate structure and semantic readiness, repairing at most three times.
- If the user requested building, hand off to execute-plan after validation. If the request was only for a plan, return the plan. Existing implementation authorization is sufficient; do not add a confirmation round.
Inputs and options
Accept a file/direct requirement or one --ado / --linear source; sources are mutually exclusive. GitHub issue URLs are valid sources. Honor --type=bug|feature|refactor|infra, --plan-path=, --parallel=auto|sequential and --design-it-twice=auto|always|never (also accept --design-twice). Auto compares designs only for significant unresolved architecture. Never skips a decision already settled. Forward recognized execution options to execute-plan; flag unknown options rather than discarding them.
Lightweight kickoff uses the same audit, readiness and dependency contract at smaller scope. Do not impose a large discovery interview on a clear, small build request.
Examples
- "Build this PRD" → audit, score, plan, validate, then implement within the request.
- "Plan this feature but don't code" → stop with the validated plan and any open decisions.
Closed decisions and open decisions
The source controls product scope. Once validated, the plan controls implementation; source-to-plan mismatches are explicit ambiguities, not silent scope expansion.
Do not
Do not assume greenfield, invent product decisions, require unsupported agent tools, or create legacy wave/lane sections alongside the task graph.
Copilot integration
Use available Copilot tools and repository instructions. This durable skill works without prompt-file discovery; the matching prompt is an optional VS Code shortcut. Do not require /fleet, /tasks or any other host-specific command when it is unavailable.