diff --git a/skills/bulk-refactor-workflow/SKILL.md b/skills/bulk-refactor-workflow/SKILL.md
index 0f20c642..fbb67826 100644
--- a/skills/bulk-refactor-workflow/SKILL.md
+++ b/skills/bulk-refactor-workflow/SKILL.md
@@ -1,65 +1,109 @@
 ---
 name: bulk-refactor-workflow
-description:
-  "Safe bulk refactors and mass edits across a repo (rename APIs, global
-  replacements, mechanical changes). Triggers: bulk refactor, mass edit, rename
-  symbol, global replace, 批量重构, 全局替换, 统一改名, 大范围修改."
+description: Use when renaming symbols, migrating APIs, or applying safe mechanical changes across many files in the repository.
 ---
 
 # Bulk Refactor Workflow（批量重构 / 大范围修改）
 
+## Overview
+
+Guide repository-wide mechanical edits without losing scope control or damaging
+unrelated work. Core principle: enumerate first, transform mechanically, verify
+in loops, and keep every batch auditable.
+
 ## When to Use
 
-- Rename API / symbol / file convention across many files
-- Mechanical refactors (imports, formatting, lint fixes, signature migrations)
-- Cross-cutting changes touching 10+ files
+- Rename APIs, symbols, file conventions, or repeated string patterns
+- Apply scripted or structural edits across many files
+- Migrate callers to a new interface where the change pattern is repeatable
 
-## Inputs（required）
+## When Not to Use
 
-- Scope：目录/文件类型/排除项（include/exclude）
-- Transformation：要做的规则（rename A→B、替换模式、接口迁移策略）
-- Constraints：是否允许行为变化？是否需要兼容期？是否允许自动格式化？
-- Verification：必须通过哪些命令/检查（最少一个）
+- One-off logic changes where no repeatable transformation exists
+- Exploratory redesign or feature work that needs new architecture decisions
+- Cleanup that is only formatting/lint alignment; use `style-cleanup` directly
 
-## Procedure（default）
+## Inputs
 
-1. **Baseline**
+- Scope: directories, file types, include/exclude rules
+- Transformation: rename, replace, codemod, scripted edit, migration strategy
+- Constraints: behavior-preserving vs. compatibility window vs. staged rollout
+- Verification commands: smallest useful loop plus final gate
 
-   - 确保工作区干净：`git status --porcelain`
-   - 跑一个基线验证（至少 build 或核心测试子集），避免“本来就坏”
+## Procedure
 
-2. **Enumerate**
+1. **Baseline and bound the scope**
 
-   - 先搜索再改：用 `rg`/`git grep` 列出全部命中
-   - 分类命中：真实调用 vs 注释/文档/样例；避免误改
+   - Record `git status --short`.
+   - Dirty worktrees are allowed.
+   - Do not revert unrelated changes.
+   - Mark the exact target file set before editing.
+   - Run a baseline verification command so existing failures are not misattributed
+     to the refactor.
 
-3. **Apply Mechanical Change**
+2. **Enumerate every intended hit**
 
-   - 优先使用确定性的机械变换（脚本/结构化编辑）而非手工逐个改
-   - 每轮改动后立即做小验证（编译/单测子集）
-   - 复杂迁移优先“两阶段”：先兼容旧接口（deprecated），再清理旧接口
+   - Search before editing with `rg`, `git grep`, or a structured query.
+   - Separate real code hits from comments, docs, generated files, and samples.
+   - Write down exclusions before running the transformation.
 
-4. **Format & Lint（按项目约定）**
+3. **Choose the safest mechanical transform**
 
-   - 仅在确认“会破坏 diff 可读性”前提下分批格式化（避免把重构和格式揉在一起）
+   - Prefer deterministic transforms: codemods, scripts, structured edits, or
+     narrow search/replace with explicit scope.
+   - If compatibility matters, use a staged migration: adapt providers first,
+     migrate callers second, remove compatibility layer last.
 
-5. **Verify & Report**
-   - 跑约定的验证命令并记录输出
-   - 汇总影响范围：改动文件数、主要改动点、潜在风险
+4. **Apply in bounded batches**
 
-## Execution Hint（optional）
+   - First, apply the transformation in bounded batches.
+   - Verify each batch before widening scope.
+   - Stop immediately if output differs from the expected hit class.
 
-如果环境支持“执行型批量处理”（例如脚本执行），优先用脚本完成批量修改，然后只把**最小 diff + 摘要**交付，避免上下文膨胀与漏改。
+5. **Verify in loops**
 
-## Output Contract（stable）
+   - Run the smallest relevant verification after each batch.
+   - Re-check the hit list after edits to confirm the intended pattern is gone and
+     unintended patterns were not introduced.
 
-- Scope：改动覆盖范围（文件/目录/语言）
-- Transformation：执行的规则（可复用）
-- Changes：关键改动摘要（按类别）
-- Verification：命令 + 证据（输出/退出码/产物）
-- Risks：高风险点与回滚建议
+6. **Finish with targeted cleanup**
+
+   - If the mechanical change leaves formatting or lint fallout, Use `style-cleanup` for the final formatting/lint pass.
+   - Keep that cleanup scoped to the refactor set unless the user approves a wider
+     pass.
+
+7. **Report**
+
+   - Summarize files touched, transform rules used, verification evidence, and any
+     residual risk.
+
+## Output Contract
+
+- `Scope:` files, directories, and exclusions actually used
+- `Transformation:` exact rule or script applied
+- `Batches:` how the work was partitioned
+- `Verification:` commands, exit status, and what each loop proved
+- `Risks:` remaining ambiguous areas, compatibility debt, or deferred cleanup
+
+## Success Criteria
+
+- Every changed file belongs to the declared scope
+- The transformation is reproducible and explainable
+- Verification passes at both batch level and final gate
+- Unrelated work in the repository is preserved
+
+## Failure Handling
+
+- If the hit list is ambiguous, stop and refine scope before editing
+- If the transform cannot be expressed mechanically, downgrade to a normal
+  refactor plan instead of pretending it is safe bulk work
+- If verification fails mid-batch, stop, inspect the smallest failing delta, and
+  correct the transform before continuing
+- If compatibility risk is higher than expected, split the migration into staged
+  commits or phases
 
 ## Guardrails
 
-- 任何“全局替换”都必须先给出命中清单与排除策略
-- 避免把行为重构与格式化/无关清理混在同一轮
+- Never run repository-wide replacement without an explicit hit review
+- Never mix unrelated cleanup into the same batch
+- Never hide a large-scope behavior change behind the label "mechanical refactor"
diff --git a/skills/commit-message/SKILL.md b/skills/commit-message/SKILL.md
index 75173bb9..f180a078 100644
--- a/skills/commit-message/SKILL.md
+++ b/skills/commit-message/SKILL.md
@@ -1,61 +1,98 @@
 ---
 name: commit-message
-description:
-  "基于 staged diff 生成符合 commit_message.md 的提交信息建议（:emoji:
-  type(scope): subject）。Triggers: commit message, 提交信息, 写提交说明,
-  生成提交信息, emoji commit, git commit."
+description: Use when the user asks for help writing a commit message, wants emoji/type(scope): subject formatting, or asks to review staged changes before committing.
 ---
 
 # Commit Message（提交信息建议器）
 
-目标：基于 `git diff --cached`（staged
-diff）生成 1–3 条提交信息建议：`:emoji: type(scope): subject`（可选 body/footer）。
+## Overview
 
-权威规范（单一真源）：
+Turn a staged diff into a commit recommendation that matches repository policy.
+Core principle: understand staged intent and commit boundaries before drafting the
+message.
+
+Authority source:
 
 - `docs/common/commit_message.md`
+- `docs/standards/playbook/docs/common/commit_message.md` (vendored playbook)
 
-## When to use
+## When to Use
 
-- 用户要“写提交信息 / 生成 commit message / 按规范写提交说明 / emoji commit”
-- 已经 `git add` 了一批改动，准备 `git commit`
+- The user asks to write or improve a commit message
+- The user wants emoji/type(scope): subject formatting
+- The user asks whether current staged changes should be split before commit
+- The user asks to review staged changes before running `git commit`
 
-## Procedure（default）
+## When Not to Use
 
-1. **收集 staged 概览（尽量小上下文）**
+- No staged changes exist and the user expects a final commit message based on the
+  actual staged diff
+- The task is writing a PR title, release note, or changelog instead of a git
+  commit message
+- The user already provided an exact commit message and only wants it executed
 
-   - `git diff --cached --name-status`
-   - `git diff --cached --stat`
-   - 必要时只看关键文件：`git diff --cached -- <path>`
+## Inputs
 
-2. **读取并遵循权威规范**
+- Current staging state (`git status --short`, `git diff --cached`)
+- Nearest applicable `commit_message.md`
+- Whether the user wants suggestions only or a command-ready final message
 
-   - 优先读取就近的
-     `commit_message.md`（见上方路径），以其中的 type/emoji/格式为准。
+## Procedure
 
-3. **生成 1 条主建议 + 2 条备选**
+1. **Baseline the repo state**
 
-   - 格式固定：`:emoji: type(scope): subject`（scope 可省略）。
-   - subject 用一句话描述“做了什么”，避免含糊词；尽量 ≤ 72 字符，不加句号。
+   - Inspect staged and unstaged state separately.
+   - If nothing is staged, stop. Explain that a staged-diff-based recommendation
+     is unavailable and ask whether files should be staged first.
+   - If only unstaged changes exist, stop before proposing a final message. Offer
+     a draft only if the user explicitly wants one.
 
-4. **判断是否建议拆分提交**
+2. **Load the authority spec**
 
-   - 当 staged 同时包含多个不相关模块/目的时：建议拆分，并给出拆分方式（按目录/功能点/风险）。
+   - Prefer the nearest `commit_message.md`.
+   - Reuse its type, emoji, scope, and body/footer conventions exactly.
 
-5. **可选：补充 body/footer（如需要）**
+3. **Summarize staged intent**
 
-   - body：说明 why/impact/verify（按规范建议换行）。
-   - footer：任务号或 `BREAKING CHANGE:`（若有）。
+   - Read the staged diff at the smallest useful granularity.
+   - Identify the main logical change, affected areas, and whether the staged set
+     mixes unrelated work.
+   - If staged changes combine unrelated goals, strongly recommend splitting the commit
+     before drafting a single message. In other words, strongly recommend splitting
+     the commit instead of hiding multiple intents behind one subject.
 
-6. **只给建议，不默认执行 `git commit`**
-   - 仅当用户明确要求时，才根据选定方案生成最终提交信息。
+4. **Draft message options**
 
-## Output contract（固定输出）
+   - Produce one recommended option and up to two alternatives.
+   - Keep the subject concrete, specific, and under the repo limit.
+   - Use body/footer only when they add real context such as motivation, impact,
+     verification, issue links, or breaking-change notes.
 
-- Detected: staged files summary（文件数 + 关键路径 + 是否可能需要拆分）
-- Proposed:
-  - Option A（recommended）：`:emoji: type(scope): subject`
-  - Option B：...
-  - Option C：...
-- Optional body/footer（如适用）
-- Notes：规范路径命中情况（哪个 `commit_message.md` 被使用）
+5. **Finalize safely**
+
+   - Make it explicit whether the output is a suggestion or a final chosen message.
+   - Do not run `git commit` unless the user explicitly asks for execution after
+     reviewing the recommendation.
+
+## Output Contract
+
+- `Detected:` staged files summary, dominant intent, and split/no-split judgment
+- `Spec:` which `commit_message.md` was used
+- `Proposed:` Option A (recommended), Option B, Option C
+- `Optional body/footer:` only when justified
+- `Notes:` risks, ambiguity, or split advice
+
+## Success Criteria
+
+- The recommendation matches the staged diff rather than guessed intent
+- The selected format complies with the nearest commit policy
+- Split advice appears when staged changes are logically mixed
+- No commit command is executed unless the user asked for it
+
+## Failure Handling
+
+- If no staged diff exists, stop and explain the limitation instead of inventing a
+  final message
+- If the staged set is too broad or ambiguous, recommend splitting and explain why
+- If no local policy file is found, state the fallback convention before drafting
+- If repo state changes during review, rerun the staged summary before finalizing
diff --git a/skills/style-cleanup/SKILL.md b/skills/style-cleanup/SKILL.md
index 64b0a8c9..a41e4ccb 100644
--- a/skills/style-cleanup/SKILL.md
+++ b/skills/style-cleanup/SKILL.md
@@ -1,94 +1,107 @@
 ---
 name: style-cleanup
-description:
-  "Clean up formatting and code style with the repo’s existing toolchain
-  (clang-format/black/isort/flake8/pre-commit/etc). Triggers: 整理代码风格,
-  格式化, format, fmt, lint fix, clang-format, black, isort."
+description: Use when the user asks to format code, fix lint issues, or align style with the repository's existing toolchain without changing behavior.
 ---
 
 # Style Cleanup Workflow（整理代码风格 / 格式化）
 
+## Overview
+
+Apply the repository's existing formatter and lint toolchain without changing
+behavior. Core principle: trust repo truth, keep scope tight, and verify with a
+rerun.
+
 ## When to Use
 
-- “整理代码风格 / 格式化 / format / fmt / lint fix”
-- 合并前做一次风格对齐（不做语义级重构）
-- 批量改动后，希望把格式化与机械性风格问题收敛到可控 diff
+- The user asks for formatting, `fmt`, `format`, or lint cleanup
+- A code change is done and needs a final style pass before review or commit
+- A mechanical change produced noisy diff that should be normalized by the repo's
+  existing toolchain
 
-## Inputs（required）
+## When Not to Use
 
-- Scope：仅本次改动文件（默认）｜全仓库｜指定目录/文件类型
-- Languages：自动识别；如为多语言仓库请确认优先级
-- Verification：至少一个可执行的验证命令（如未知，先问/再推断）
+- This is not for semantic refactors or behavior changes
+- This is not for introducing a new formatter or lint configuration
+- This is not for full-repo reformatting unless the user explicitly requests that
+  scope
 
-## Procedure（default）
+## Inputs
 
-1. **Baseline**
+- Target scope: staged files, unstaged files, explicit file list, or explicit
+  directory
+- Relevant languages in scope
+- Verification commands available in the repository
 
-   - 记录当前状态：`git status --porcelain`
-   - 明确范围（默认只处理变更文件）：
-     - staged：`git diff --name-only --cached`
-     - unstaged：`git diff --name-only`
-     - untracked：`git ls-files -o --exclude-standard`
+## Procedure
 
-2. **Detect Toolchain（prefer repo truth）**
+1. **Baseline the scope**
 
-   - 优先用仓库既有入口脚本 / 配置：
-     - JS/TS：`package.json`
-       scripts（`format`/`lint`/`lint:fix`）、prettier/biome/eslint 配置
-     - Python：`pyproject.toml` / `.flake8` / `.pylintrc` /
-       `.pre-commit-config.yaml`
-     - C/C++：`.clang-format`（唯一真相），可选 `.clang-tidy`
-     - Shell：`shfmt`/`shellcheck`（若仓库已使用）
-     - Markdown：prettier/markdownlint（仅在仓库已固定时使用）
-   - 禁止默认“引入新 formatter/linter 配置”；缺配置时只做最小手工调整，并先确认是否允许落地配置文件。
+   - Record `git status --short`.
+   - Default to changed files only unless the user requested broader scope.
+   - Resolve the target file set before running any formatter.
 
-3. **Apply（format first, then lint）**
+2. **Detect the repo toolchain**
 
-   - 先 formatter（会改文件），再 lint（检查），再 lint
-     --fix（如有），最后再跑一次 check 确认干净。
-   - 默认只处理目标文件集合；避免全仓库 reformat（除非用户明确要求）。
-   - 典型命令（按仓库实际替换）：
-     - C++：`clang-format -i <files...>`；CI 校验：`clang-format --dry-run --Werror <files...>`
-     - Python：`black <files...>` + `isort <files...>`；或
-       `pre-commit run --files <files...>`
-     - JS/TS：`npm run format -- <files...>` / `pnpm ...` /
-       `npx prettier -w <files...>`（以项目脚本为准）
+   - Prefer repository scripts and checked-in config over ad hoc commands.
+   - Use the repo's existing formatter/linter stack:
+     - JS/TS: `package.json`, prettier, biome, eslint
+     - Python: `pyproject.toml`, `isort`, `black`, `ruff`, `flake8`,
+       `pre-commit`
+     - C/C++: `.clang-format`, optional `.clang-tidy`
+     - Shell: `shfmt`, `shellcheck` if already present
+     - Markdown: prettier/markdownlint only if the repo already uses them
 
-4. **Guardrails**
+3. **Run cleanup in a fixed loop**
 
-   - 只做风格与格式：不改变行为、不改 public API、不做重构。
-   - 如格式化导致 diff 暴涨（文件数/行数过大）：先停下，给出原因与两种方案让用户选：
-     1. 仅格式化本次改动文件（推荐默认）
-     2. 全仓库统一格式（通常需要单独 PR/提交）
+   - Use `formatter -> lint/check -> lint --fix -> final check`.
+   - Prefer repository entrypoints over raw tool invocations.
+   - Keep execution scoped to the chosen files whenever the toolchain supports it.
+
+4. **Control blast radius**
+
+   - If the formatter expands the diff far beyond the requested scope, stop and
+     explain the tradeoff.
+   - Offer a scoped pass vs. a full-repo normalization pass instead of silently
+     widening scope.
 
 5. **Verify**
-   - 跑最小验证命令（仓库已有命令优先）。
-   - 若无法运行（缺环境/缺权限/缺依赖）：说明原因，并给出替代验证（例如 formatter 二次运行无 diff、lint 输出为 0）。
 
-## Playbook as Authority（如果项目 vendoring 了本 Playbook）
+   - Re-run the lightweight checks after fixes.
+   - Confirm that the second formatter run produces no additional diff.
+   - If the repo has a canonical lint or test command for style verification, run
+     that command and report the result.
 
-当目标仓库包含 `docs/standards/playbook/docs/`（或直接包含
-`docs/tsl|cpp|python/...`），风格决策参考：
+## Playbook as Authority
 
-- TSL：`docs/tsl/code_style.md`、`docs/tsl/naming.md`、`docs/tsl/toolchain.md`
-- C++：`docs/cpp/code_style.md`、`docs/cpp/naming.md`、`docs/cpp/toolchain.md`
-- Python：`docs/python/style_guide.md`、`docs/python/tooling.md`、`docs/python/configuration.md`
+When the target repo vendors this playbook, prefer these references for style
+judgment:
 
-## Output Contract（stable）
+- TSL: `docs/tsl/code_style.md`, `docs/tsl/naming.md`, `docs/tsl/toolchain.md`
+- C++: `docs/cpp/code_style.md`, `docs/cpp/naming.md`, `docs/cpp/toolchain.md`
+- Python: `docs/python/style_guide.md`, `docs/python/tooling.md`,
+  `docs/python/configuration.md`
 
-- Scope：实际处理范围（文件/目录/语言）
-- Toolchain：使用了哪些工具与配置依据
-- Commands：实际执行命令（按顺序）
-- Changes：修改文件列表 + 改动规模概览
-- Remaining：仍未修复的问题（分类：formatter / lint / style）+ 下一步建议
+## Output Contract
+
+- `Scope:` actual files/directories/languages processed
+- `Toolchain:` repo configs and commands used
+- `Commands:` execution order
+- `Changes:` modified files plus diff-size summary
+- `Remaining:` unresolved lint/style issues plus why they remain
 
 ## Success Criteria
 
-- formatter 二次运行无新增 diff
-- lint/检查命令通过（或仅剩已确认的例外）
-- 未引入语义变更（仅格式/风格）
+- Style cleanup stays within the requested or agreed scope
+- The second formatter run produces no additional diff
+- Verification commands pass, or any remaining exception is explicitly explained
+- No semantic behavior change is introduced
 
 ## Failure Handling
 
-- 工具缺失：优先提示安装方式或替代命令；无法解决则退回“最小手工风格修复 + 明确未覆盖项”
-- 规则冲突（如 black vs flake8）：以仓库配置为准；必要时调整例外配置但需先确认
+- If the repo lacks a formatter/linter, say so and fall back to minimal manual
+  cleanup only with clear limits
+- If tools are missing locally, report the missing dependency and stop before
+  inventing a new toolchain
+- If two configured tools conflict, follow checked-in repo config and surface the
+  conflict instead of papering over it
+- If cleanup would require wider scope than requested, stop and ask for approval
diff --git a/tests/README.md b/tests/README.md
index 85f10b0e..b0dc7777 100644
--- a/tests/README.md
+++ b/tests/README.md
@@ -10,6 +10,7 @@ tests/
 ├── cli/                            # Python CLI 测试（unittest）
 │   └── test_playbook_cli.py        # playbook.py 基础功能测试
 ├── test_format_md_action.py        # format_md 动作测试
+├── test_firstparty_skills_quality.py # first-party skills 元数据与结构质量测试
 ├── test_gitattributes_modes.py     # gitattr_mode 行为测试
 ├── test_no_backup_flags.py         # no_backup 行为测试
 ├── test_sync_directory_actions.py  # sync_memory_bank/sync_prompts 行为测试
diff --git a/tests/test_firstparty_skills_quality.py b/tests/test_firstparty_skills_quality.py
new file mode 100644
index 00000000..40bdff1c
--- /dev/null
+++ b/tests/test_firstparty_skills_quality.py
@@ -0,0 +1,104 @@
+import re
+import unittest
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[1]
+SKILLS_ROOT = ROOT / "skills"
+FIRST_PARTY_SKILLS = {
+    "commit-message": SKILLS_ROOT / "commit-message" / "SKILL.md",
+    "style-cleanup": SKILLS_ROOT / "style-cleanup" / "SKILL.md",
+    "bulk-refactor-workflow": SKILLS_ROOT / "bulk-refactor-workflow" / "SKILL.md",
+}
+
+
+def read_text(path: Path) -> str:
+    return path.read_text(encoding="utf-8")
+
+
+def normalize_space(text: str) -> str:
+    return " ".join(text.split())
+
+
+def parse_frontmatter(text: str) -> dict[str, str]:
+    match = re.match(r"^---\n(.*?)\n---\n", text, re.DOTALL)
+    if match is None:
+        raise AssertionError("missing YAML frontmatter")
+    block = match.group(1)
+    data: dict[str, str] = {}
+    current_key: str | None = None
+    current_value: list[str] = []
+    for raw_line in block.splitlines():
+        if raw_line.startswith("  ") and current_key is not None:
+            current_value.append(raw_line.strip())
+            continue
+        if current_key is not None:
+            data[current_key] = " ".join(current_value).strip().strip('"')
+            current_key = None
+            current_value = []
+        key, value = raw_line.split(":", 1)
+        current_key = key.strip()
+        current_value = [value.strip()]
+    if current_key is not None:
+        data[current_key] = " ".join(current_value).strip().strip('"')
+    return data
+
+
+class FirstPartySkillsQualityTests(unittest.TestCase):
+    def test_first_party_skill_frontmatter_is_minimal_and_named_consistently(self):
+        for name, path in FIRST_PARTY_SKILLS.items():
+            with self.subTest(skill=name):
+                frontmatter = parse_frontmatter(read_text(path))
+                self.assertEqual(set(frontmatter), {"name", "description"})
+                self.assertEqual(frontmatter["name"], name)
+                self.assertRegex(frontmatter["name"], r"^[a-z0-9-]+$")
+
+    def test_first_party_skill_descriptions_are_trigger_focused(self):
+        for name, path in FIRST_PARTY_SKILLS.items():
+            with self.subTest(skill=name):
+                description = parse_frontmatter(read_text(path))["description"]
+                self.assertTrue(description.startswith("Use when"))
+                self.assertLessEqual(len(description), 500)
+                self.assertNotIn("Triggers:", description)
+
+    def test_first_party_skills_have_required_sections(self):
+        required_sections = (
+            "## Overview",
+            "## When to Use",
+            "## When Not to Use",
+            "## Inputs",
+            "## Procedure",
+            "## Output Contract",
+            "## Success Criteria",
+            "## Failure Handling",
+        )
+        for name, path in FIRST_PARTY_SKILLS.items():
+            text = read_text(path)
+            with self.subTest(skill=name):
+                for section in required_sections:
+                    self.assertIn(section, text)
+
+    def test_commit_message_skill_handles_missing_or_mixed_staging_states(self):
+        text = normalize_space(read_text(FIRST_PARTY_SKILLS["commit-message"]))
+        self.assertIn("If nothing is staged", text)
+        self.assertIn("If only unstaged changes exist", text)
+        self.assertIn("strongly recommend splitting the commit", text)
+        self.assertIn("Do not run `git commit`", text)
+
+    def test_style_cleanup_skill_has_clear_non_goals_and_verification_loop(self):
+        text = normalize_space(read_text(FIRST_PARTY_SKILLS["style-cleanup"]))
+        self.assertIn("not for semantic refactors", text)
+        self.assertIn("not for introducing a new formatter or lint configuration", text)
+        self.assertIn("formatter -> lint/check -> lint --fix -> final check", text)
+        self.assertIn("second formatter run produces no additional diff", text)
+
+    def test_bulk_refactor_skill_is_dirty_aware_and_delegates_final_cleanup(self):
+        text = normalize_space(read_text(FIRST_PARTY_SKILLS["bulk-refactor-workflow"]))
+        self.assertIn("Dirty worktrees are allowed", text)
+        self.assertIn("Do not revert unrelated changes", text)
+        self.assertIn("Use `style-cleanup` for the final formatting/lint pass", text)
+        self.assertIn("apply the transformation in bounded batches", text)
+
+
+if __name__ == "__main__":
+    unittest.main()