Agent skill wrapper for the Claude command
.
When the original command text references
,
, or named arguments, map them from the user's current request.
Execute the implementation plan for a feature by working through
phase by phase, using TDD where applicable.
-
Locate artifacts:
- If input names a feature, directory, or phase number, parse it.
- Load required: , . Load optional: , , .
- Load
.specify/memory/constitution.md
.
- Set to the feature directory.
-
Check checklists: Scan
for any uncompleted items (
). If found, report them and ask: "There are incomplete checklist items. Proceed anyway?" Stop if the user declines.
-
Determine starting phase:
- If a phase number was provided in input, start there.
- Otherwise start from Phase 1.
- Skip phases where all tasks are already marked .
-
Execute each phase:
- Announce the phase name and task count before starting.
- For tasks marked (parallelizable): execute them concurrently.
- For tasks without : execute in listed order.
- For each task:
- Write tests first (TDD), then implement the minimum code to satisfy the task.
- Mark the task in immediately after completion.
- Run available test and lint commands to verify no regressions.
- If a non-parallelizable task fails, stop and report. Do not proceed to the next task until resolved.
-
Validate completion after all phases:
- All tasks in are marked .
- All functional requirements from are addressed.
- Tests pass and no new lint errors.
-
Report: Tasks completed, any skipped tasks with reasons, test results summary.
Next step: Run
to track remaining work as GitHub issues, or commit and open a PR.