SKILL.md
Full skill instructions
Implement
Execute the current implementation plan. Follow the plan step by step, validate as you go, keep the repo green.
Required Skills
spec-driven-development(state/task management, lane detection)researchagent-browser(required when implementing frontend work to validate behavior, semantics, and computed CSS)
Inputs
[!IMPORTANT] Resolve the change set by running
ls changes/ | grep -v archive/. If exactly one directory exists, use it. Only prompt the user when multiple change sets are present.
Instructions
-
Load
spec-driven-developmentskill and read state fromchanges/<name>/state.toml. Apply state entry check per skill guidelines. -
Read tasks from
changes/<name>/tasks.tomlif full lane. -
Determine lane and load plan:
- Full lane: Identify current task (in_progress status). Read corresponding plan from
changes/<name>/plans/ - Vibe/Bug lane: Read
changes/<name>/plan.md(single combined plan)
- Full lane: Identify current task (in_progress status). Read corresponding plan from
-
Full lane phase/task entry checks:
- Confirm the active phase is
implement - Confirm exactly one task is
in_progressbefore execution
- Confirm the active phase is
-
Execute the plan step by step:
- Follow steps exactly as written
- Validate after significant changes
- Keep repo green
- Use
researchskill for unexpected code structure or integration questions - For any frontend implementation (UI, styling, layout, interaction, accessibility), use
agent-browserskill to validate the implemented result in a browser - During frontend validation, explicitly verify semantic HTML correctness and that computed CSS matches the intended design/reference used during implementation
- Document any deviations from plan
-
Handle issues:
- Minor adjustments: Proceed and document deviation
- Major issues: Stop and discuss with user
- Spec issues (full lane): Flag for reconciliation
-
Run validation steps from plan, verify acceptance criteria, ensure tests pass.
-
Full lane completion:
- Mark current task complete with
ae sdd task complete [name](orae sdd task complete --next [name]when explicitly chaining tasks) - Do not manually edit
tasks.tomltask status when CLI task commands are available - After task completion, suggest
ae-sdd-next <name>:- If tasks remain,
ae-sdd-nextshould deterministically loopimplement -> plan - If tasks are complete,
ae-sdd-nextshould advanceimplement -> reconcile
- If tasks remain,
- Do not update phase status directly in this command
- Mark current task complete with
-
Vibe/Bug lane completion:
- Discuss next steps with user
- If keeping work: do not update phase status; suggest
ae-sdd-next <name>(optional) - If throwing away: done, no state update needed
Examples
Full lane implementing a task:
Input: None (change: "password-reset")
Output: "Loading plan 01.md to implement password validator changes."
Follows steps: update validator.ts, add reset logic, update tests.
Validation: All tests pass.
User: "Looks good."
Output: "Marked task 1 complete. Three tasks remaining—run ae-sdd-next to loop back to plan for the next task."
Vibe lane quick fix:
Input: "bug-fix" (user has context with plan.md)
Output: "Following plan.md to patch router/routes.ts."
Implementation complete, tests pass.
User: "Great, keep this work."
Output: "Implementation looks complete. If you want to advance phase, run ae-sdd-next."
