一键导入
refine-skill
Review and improve an existing Claude Code skill. Use when asked to refine a skill, improve a skill, review a skill, or clean up a skill.
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。
菜单
Review and improve an existing Claude Code skill. Use when asked to refine a skill, improve a skill, review a skill, or clean up a skill.
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
Disciplined diagnosis loop for hard bugs and performance regressions. Reproduce → minimise → hypothesise → instrument → fix → regression-test. Use when user says "diagnose this" / "debug this", reports a bug, says something is broken/throwing/failing, or describes a performance regression.
Pre-push Claude-driven code review loop with auto-fix. Runs parallel review agents (CLAUDE.md, bugs, git history, code comments, security, test coverage, and — on substantial diffs — structural simplification and observability coverage; plus, in cycle 1 on feature work, an intent-reconciliation agent that models what the PR should do from its purpose and reconciles that against the code to catch omission bugs) per cycle, scores each finding 0-100 with an independent Haiku rater, auto-fixes ≥80 findings, asks about 50-79 in a single batch per cycle, and stores per-repo learnings in .git/info/. Each cycle first runs the static-analysis skill (a deterministic linter/security/secret tool pass with autofix) so the LLM agents can focus on what tools can't catch. On clean exit, a final manual-testing evidence gate verifies the PR carries proof the changed functionality actually works — and produces that proof itself (local dev servers + playwright + surge-image-upload) when missing — reconciles the PR description to
Run every applicable static-analysis tool on a repo — detect languages/configs, run the curated CodeRabbit-weighted analyzer set (installed or ephemerally via npx/uvx), and write results to .static-analysis/. Use when asked to run static analysis, lint the whole repo, run all linters/analyzers, do a code-quality/security scan, or check a repo before pushing. Report-only by default; --fix opts into safe autofixers. Subagent-safe (never blocks).
Drive a pull request to a mergeable state: no conflicts with the base branch, green CI, and a clean review. Integrates the base branch and resolves conflicts, delegates PR feedback resolution (human + bot/CodeRabbit) to the resolve-feedback skill, runs the local review loop, fixes failing CI (via the fix-ci skill), and pushes. Waits once for CodeRabbit's initial automatic review to land and resolves its feedback, but does not re-request reviews or wait for CodeRabbit again after later pushes (CodeRabbit reviews each PR once, on open). Use when the user wants to drive a PR forward / get it to green / make it mergeable, resolve PR comments or review feedback, fix CI, or resolve merge conflicts — even if there is no feedback yet. Drives to mergeable but does NOT merge.
Retrieve, classify, and resolve PR review feedback — inline review threads, review summaries, and generic PR comments, from humans and bots. Fixes valid feedback (committing locally), posts justifications for invalid feedback, files follow-ups for out-of-scope items, and — critically — literally marks each item resolved/dismissed on GitHub (a reply is NOT a resolution). Supports an interactive mode (prompts on human feedback) and a non-interactive/batch mode (auto-applies bot/procedural feedback, defers human feedback to the caller). Use when the user wants to resolve PR feedback / review comments / review threads, address reviewer comments, or 'resolve feedback'. This is the feedback engine that the drive-pr and pr-triage skills delegate to; it does not push, run the review loop, fix CI, or integrate the base branch — that is the caller's job.
Audit and sync all pending changes across the dotfiles and dotprivate bare repos. Use when asked to check for uncommitted changes, sync dotfiles, or commit and push dotfile updates.
| name | refine-skill |
| description | Review and improve an existing Claude Code skill. Use when asked to refine a skill, improve a skill, review a skill, or clean up a skill. |
You are a skill reviewer. Your role is to audit an existing skill and propose concrete improvements based on the principles below.
SKILL.md is hot-path only. Every line in SKILL.md is loaded into context on every invocation. Only include information that is useful on the majority of invocations. Move situational knowledge to reference files.
Scripts over prose for repeatable procedures. If a workflow step should be done the same way every time, put it in a script. SKILL.md should call the script, not spell out the commands.
Reference files for situational knowledge. Information that is pertinent to the skill but only needed sometimes belongs in separate files that the LLM can read on demand. Link to them from SKILL.md.
Correct over cautious. If past experience revealed a wrong assumption baked into the skill, fix it. Don't add hedging language — update the guidance to be accurate.
Delete stale content. If a tip, workaround, or command is no longer needed, remove it. Don't comment it out or add "deprecated" notes.
If not specified, ask the user which skill to review. Read the full skill directory:
ls -la ~/.claude/skills/{skill-name}/
# or for project skills:
ls -la ${PWD}/.claude/skills/{skill-name}/
Read SKILL.md and all supporting files.
For each section of SKILL.md, ask:
Review recent invocations or the user's description of what went wrong. Look for:
Present changes organized by type:
Get user approval before making changes.
Make the approved changes. After writing:
--help output — only document non-obvious usage