executing-plans

name	executing-plans
description	Use after an approved implementation plan exists and the user wants execution. Runs the plan task-by-task with checkpoints, optional project-agent loop, tests, reviews, and final verification without expanding scope.

Executing Plans

Reference: https://github.com/obra/superpowers

Use this skill only after skills/writing-plans/SKILL.md has produced an approved plan. The plan is the source of truth; do not add scope unless the user approves a plan update.

Preconditions

• An implementation plan exists and has been approved by the user.
• If work should be isolated, use skills/using-git-worktrees/SKILL.md before editing.
• Read the plan once, extract tasks, files, tests, and acceptance criteria.
• Confirm whether execution is inline or project-agent loop.

Inline Execution

Use inline execution for small, clear, tightly coupled work.

For each task:

1. Mark exactly one task in progress.
2. Perform the next unchecked step only.
3. Run the command specified by the plan.
4. If a test or command fails unexpectedly, pause and use skills/systematic-debugging/SKILL.md.
5. Do a requirement/spec compliance check for the task.
6. Do a code quality check for the task.
7. Mark the task complete only after checks pass.

Project-Agent Loop

Use this for substantial plans with independent tasks.

For each task:

1. Dispatch one fresh implementation agent with only:
   • the approved spec summary
   • the exact task section
   • relevant file paths
   • expected tests and acceptance criteria
2. Tell the agent it is not alone in the codebase and must not revert others' changes.
3. After implementation, run requirement/spec compliance review.
4. Run code quality review with agents/code-reviewer.md; add security or database review only when the touched area warrants it.
5. Send findings back to the same implementer until resolved.
6. Mark the task complete only after tests and reviews pass.

Subagents must return only conclusion, evidence paths, risks, and next steps. Do not accept large raw logs as final output.

Completion Loop

After all tasks are complete:

1. Apply skills/verification-before-completion/SKILL.md: identify the verification evidence required before any completion claim.
2. Run /verify or the plan's equivalent verification commands to produce that fresh evidence.
3. If verification fails, use skills/systematic-debugging/SKILL.md for each failure class before changing code.
4. Run a final review over the full diff.
5. Summarize files changed, tests run, skipped checks, remaining risk, and whether the work is ready for /pr.

Boundaries

• Do not commit, push, create PRs, delete branches, or remove worktrees unless the user asked for that action.
• Do not skip failed tests because later tasks might fix them; record the failure and debug now unless the plan explicitly says the failure is expected.
• Do not mark multiple tasks complete at the end; update status incrementally.