菜单
Captures high-value learnings from a skill execution and applies targeted improvements to the SKILL.md. Generic: works for any skill. Invoke right after running a skill when steps diverged from the documented flow, new techniques were discovered, or undocumented edge cases appeared. Synthesizes from conversation context automatically.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Diagnosis loop for hard bugs and performance regressions. Use when the user says "diagnose"/"debug this", or reports something broken/throwing/failing/slow.
A relentless interview to sharpen a plan or design.
Grill the user relentlessly about a plan or design. Use when the user wants to stress-test a plan before building, or uses any 'grill' trigger phrases.
Reviews text for weasel words, vague language, and corporate jargon using Amazon's writing principles. Returns flagged issues and a polished version. Use when drafting Slack updates, PR descriptions, Jira tickets, or any professional communication that needs to be direct and concrete.
Audits and rewrites CLAUDE.md files. Use when the user asks to review, optimize, rewrite, standardize, or improve a CLAUDE.md file. Triggers include "audit claude.md", "otimiza o claude.md", "melhora o claude.md", "padroniza claude.md", "review CLAUDE.md". Also use when someone says "clean up my docs", "my claude context is messy", or asks to "initialize" context files in a project.
Pre-close session review. Scans for loose ends before ending a conversation — documentation gaps, unsaved memory, open checklist items, uncommitted git changes, and notification gaps. Reports findings grouped by category; never acts without confirmation. Invoke at the end of any work session before signing off.
| name | skill-refine |
| description | Captures high-value learnings from a skill execution and applies targeted improvements to the SKILL.md. Generic: works for any skill. Invoke right after running a skill when steps diverged from the documented flow, new techniques were discovered, or undocumented edge cases appeared. Synthesizes from conversation context automatically. |
Refine a skill's SKILL.md based on what actually happened this session. Only works with same-session context — the skill must have been executed in the current conversation.
If the skill name is not passed as an argument, infer it from the most recently executed skill in the conversation. If ambiguous, ask: "Confirming: refining {name}, correct?"
Read ~/.claude/skills/{name}/SKILL.md. Since these are symlinks into the dot-claude-code repo, editing the live path updates the canonical file — verify with readlink ~/.claude/skills/{name} if unsure.
Scan the conversation for what diverged or was discovered. Focus on:
Filter ruthlessly. Only capture learnings that are non-obvious, not already implied by existing text, and likely to recur. If a learning is ambiguous, ask before adding it.
For each learning, identify where it belongs:
Show a concise before/after for each change. Confirm with the user before applying.
Surgical edits only. Don't rewrite sections that don't need changing. Don't touch frontmatter, section titles, or formatting conventions unless the learning requires it.
If doing a structural refactor, verify explicitly that these survive: priority/ordering rules, negative constraints (never/always/do not), output shape rules, and decision tree logic. None of these are derivable from structure — they must be explicitly present in the new version.
Apply confirmed edits with the Edit tool at ~/.claude/skills/{name}/SKILL.md. Then: "Run /commit to version the changes."
