autopilot_skill

Overview

This skill runs PRD-driven autonomous execution: provide a task and it plans, does, and reviews in a loop until acceptance criteria are met or a stop condition is hit. It targets single-story tasks (feature or bug) and produces machine-readable outputs like prd.json, a progress log, and a completion journal. Triggers include "autopilot", "autonomous", "get it done", "finish this", and "ship it".

How this skill works

The skill auto-generates a PRD with testable acceptance criteria and priority, then iterates through a Plan → Do → Review loop up to a configurable max (default 5). The Plan step creates granular tasks and a roadmap; Do executes and commits minimal changes; Review validates acceptance criteria and either marks COMPLETE or emits a REVIEW_FAILED reason to trigger another iteration. It enforces one task at a time, checks project map files before edits, and logs progress and outcomes.

When to use it

• You want hands-off implementation of a single feature described by acceptance criteria.
• You need a bug fixed with a reproducible test and hands-off validation.
• You want automated, repeatable execution that loops until tests pass or max iterations are reached.
• You have a task that can be verified with clear pass/fail acceptance criteria.
• You prefer minimal, incremental changes and audit-friendly execution logs.

Best practices

• Provide a clear, concise description with explicit acceptance criteria or let the skill draft them.
• Limit requests to a single story (one feature or bug) for predictable iteration and results.
• Keep MAP.md or equivalent project map up to date so the planner can find files safely.
• Enable CI/tests or provide reproducible verification steps so the reviewer can validate changes.
• Set a reasonable max iterations and use verbose logging for debugging complex tasks.

Example use cases

• Add a dark mode toggle with UI tests and build verification.
• Fix a browser-specific login bug and add a regression test.
• Implement a small API endpoint with documented acceptance criteria and automated tests.
• Refactor a single module to match project patterns and ensure build passes.
• Pick the next item from TODO.md and complete it end-to-end with validation.