菜单
Full codebase sweep scoring every file 0-10 on polish potential, refining files scoring 5+ via parallel agents. Unlike /simplify (recently changed files only), /polish sweeps the entire codebase.
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
Search indexed reference codebases (Convex, Expo, Next.js, Better Auth, Remotion, etc) via the qmd MCP query tool. Use when looking up framework APIs, finding code examples in third-party repos, or answering questions about external libraries that aren't in the current working directory.
Adds a reference collection to qmd. Accepts a GitHub URL, owner/repo shorthand, or a local directory path. Auto-detects file types, sets ignore globs, indexes with AST chunking, embeds, and verifies. Use when adding a new reference codebase or local notes folder for search.
Convert a video or screen recording to GIF. Handles HDR recordings, macOS filenames with spaces, and ffmpeg two-pass palette compression. Do NOT use for frame extraction (use /frames instead).
Extract frames from a video file or screen recording (.mov, .mp4, .webm, .avi, .mkv) so Claude can view, review, or analyze it. Do NOT use for GIF conversion (use /gif instead).
End session, save progress, create handoff. Triggers: wrap up, done for today, end session, save my work, save progress, close out. Runs health checks and captures resume point.
Hydrate session context from last handoff. Triggers: resume, pick up where I left off, continue from last session, what was I working on, session status, what's the state. Use for deep hydration beyond the auto-injected summary.
| name | polish |
| description | Full codebase sweep scoring every file 0-10 on polish potential, refining files scoring 5+ via parallel agents. Unlike /simplify (recently changed files only), /polish sweeps the entire codebase. |
ultrathink
You coordinate a full codebase polish. You score every source file 0-10 on polish potential and dispatch parallel background agents to refine files scoring 5+.
You are a coordinator. You read and analyze files, create tasks, and spawn worker agents. You do NOT edit files directly.
$ARGUMENTS containing --dry-run: analyze and report only, no modifications$ARGUMENTS containing a path: scope to that directory/fileGlob all source files using parallel patterns:
**/*.ts, **/*.tsx, **/*.js, **/*.jsx**/*.py, **/*.go, **/*.rs**/*.vue, **/*.svelteExclude: node_modules/**, dist/**, build/**, .next/**, coverage/**, *.min.*, *.d.ts, _generated/**, .git/**
If path argument provided, scope discovery to that path.
Read ${CLAUDE_SKILL_DIR}/references/workflow.md for the analysis, work queue, parallel agents, and verification workflow.
*.d.ts, _generated/, codegen output) get picked up by glob patterns. The exclusion list must be checked before scoring or agents will "polish" auto-generated code.