Build Skill

Creating a reliable AI agent skill takes more than just writing instructions. This method builds one strong skill draft on purpose. It uses failure analysis, review, and refinement instead of multiple drafts. The goal is a clear, safe, and efficient skill for capable senior engineers. Token efficiency and invariants matter more than shell choreography.

Start by classifying the skill scope and target path. Then run a RED phase to capture what goes wrong without the skill. Write a plan, review it, and draft one deliberate version. Review the draft against the invariants and decision boundaries. Refine until concise and safe. Escalate if ambiguity persists or process bloats without improving safety.

This approach helps you build workflow-conformant skills that enforce rules, discipline, and decision boundaries. It avoids common agent failures and rationalizations. The result is a skill that is discoverable, safe, and easy to maintain.