菜单
Analyze a codebase for quality issues across code smells, clean code, separation of concerns, modularity, and API-first design, then generate a prioritized refactoring report and gated execution roadmap (characterization safety net, commit discipline, delete packet, close-out delta).
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/Chaddacus/chad_claude_configs --skill refactor命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | refactor |
| description | Analyze a codebase for quality issues across code smells, clean code, separation of concerns, modularity, and API-first design, then generate a prioritized refactoring report and gated execution roadmap (characterization safety net, commit discipline, delete packet, close-out delta). |
| context | fork |
| policy_doc_kind | skill |
| classification | canonical |
| canonical_owner | self |
| authority_level | procedural |
| in_verifier_scope | true |
| lexical_guard_profile | stale_names,destructive_rollback,branch_policy_live |
This skill owns the refactor workflow and its gates. Global policy owns runtime routing, git policy, review requirements, and delivery constraints; ~/.claude/standards/testing-standard.md owns test-breadth definitions; the planner agent owns DAG mechanics.
/refactor /path/to/project
/refactor /path/to/project --focus api,modularity
/refactor /path/to/project --report-only
/refactor /path/to/project --depth shallow
| Flag | Effect |
|---|---|
(none) | Full analysis and roadmap generation |
--focus X,Y | Limit the dimensions analyzed |
--report-only | Produce findings without build-spec generation |
--depth shallow | Skip deep architecture tooling |
Refactor execution requires a recorded forcing reason — one of:
"Looks legacy" or style preference fails the gate. --report-only analysis is exempt; build-spec generation and all execution require the forcing reason recorded in the report header.
For each finding, report:
Dimensions:
Gather supporting metrics such as:
any countRecord each metric with the exact command used, as the baseline table in the report. The same commands are re-run verbatim at track close (step 8).
Write:
Before the first transform lands:
test-strategist proactively. On refactor tracks this inverts its reactive QA-stage role: the safety net is a precondition, not a gap-closure afterthought.acceptance_check is the test command. Every transform packet must list its seam's pinning packet in blocked_by; a transform packet with no pinning ancestor for its seam is a planning defect.~/.claude/standards/testing-standard.md: a refactor on an uncovered surface is never smoke-eligible.Unless --report-only, emit build specs shaped as planner packets — id, goal, owned_files, blocked_by, lane, acceptance_check (a shell command whose exit 0 proves the slice's definition of done) — plus two refactor-specific properties:
mechanical: true|false — mechanical transforms (renames, import moves, API call-site migrations, format conversions) are codemod work: ast-grep (polyglot), jscodeshift (JS/TS), OpenRewrite (JVM). A hand-edited mechanical transform across more than 3 files is a reviewer reject.acceptance_check is a grep-to-zero command, e.g. ! rg -q 'legacyHandler|LEGACY_ROUTE_FLAG' src/. The track cannot close while the delete packet is open. A strangler path that never strangles is two systems: any surviving symbol must be named in the close record with a follow-up owner.A slice is exactly one of:
Never both in one diff. A mixed diff is a reviewer reject reason.
At track close, re-run the exact baseline commands from step 3 and report the before/after table. A close record without the delta table is weak closure under global support-confidence policy.