My Agent Skills All Said How to Start. None Said When to Stop.

Описание и применение

В файлах навыков агента отсутствует информация о завершении работы и о том, как распознать ошибки. Добавлены три раздела – об обоснованиях, сигналах ошибок и проверке завершения, что позволяет агенту и человеку определить, когда работа действительно завершена.

Задача

Agent skill files focus only on how to start tasks, lacking guidance on when a task is truly finished or how to detect wrong turns.

Результат

Skills now include structured completion criteria, making it easier for agents and humans to determine when work is truly finished and to spot missteps early.

Рабочий процесс

• 1. Identify high‑risk skills (e.g., deploy, code‑review, debugging).
• 2. Append a Common Rationalizations table to counter typical shortcuts.
• 3. Add a Red Flags list of observable indicators of missteps.
• 4. Include a Verification checklist of exit conditions to run before marking work done.
• 5. Place the Verification section at the end of the skill file for auditability.
• 6. Update tooling references and conventions (e.g., companion tools, reference files).
• 7. Test the updated skills by running them in the agent and verifying the new sections trigger correctly.

Доказательства

Claude Toolkit 2.5.0 adds three sections — Common Rationalizations, Red Flags, Verification — to the nine skills where being wrong is expensive.

Источник

A checklist under a ## Verification heading is a thing you can be asked to run — "run the verification section" — and the answer is auditable, box by box.

Источник