Three Claude Code skills that codify the article's loop rules so the agent invokes them on the right turn. Each skill is a single SKILL.md with frontmatter (name, description) plus instructions in the body.
Copy any skill folder into one of these locations:
- Project:
.claude/skills/<skill-name>/SKILL.md - User:
~/.claude/skills/<skill-name>/SKILL.md
Claude Code discovers skills by directory and uses the description frontmatter field to decide when to invoke them. Tweak the descriptions for your own project's vocabulary so the right skill fires on the right turn.
| Skill | When it fires | Codifies |
|---|---|---|
verify-before-done |
About to declare a task complete | §10: never let "done" mean "the model says it is done" |
read-before-edit |
About to edit a file in unfamiliar code | §3: most engineering is reading |
loop-failure-triage |
A tool call just failed | §5: errors are not edge cases, they are the work |
The article identifies three places where loops most commonly fail: the moment of "done" (§10), the moment of edit (§3), and the moment after a failed tool call (§5). One skill per failure surface, deliberately scoped, so the agent has clear, invokable rules at each decision point instead of a single long instruction file that gets ignored.