llm-wiki-lint

$npx mdskill add KerberosClaw/kc_ai_skills/llm-wiki-lint

Identifies the path of a Karpathy LLM Wiki repo

  • Detects and returns the provided path for linting.
  • Depends on user input via `$ARGUMENTS` to identify the repository.
  • Uses pattern matching to locate the correct directory structure (raw/ + wiki/).
  • Delivers the identified path directly to the user.

SKILL.md

.github/skills/llm-wiki-lintView on GitHub ↗
---
name: llm-wiki-lint
description: "Use when the user wants to lint a repo following the Karpathy LLM Wiki pattern (raw/ + wiki/ + SCHEMA.md / index.md / log.md). The skill detects path, scans wiki pages + schema layer, reports frontmatter gaps, broken in-body links, source traceability breaks, stale claims, orphan pages, index synopsis contradictions, and SCHEMA-vs-reality drift. Phase 1 is read-only report; Phase 2 (fix + log.md LINT entry) only runs after the user explicitly approves."
version: 0.3.1
status: mvp
triggers:
  - "/llm-wiki-lint"
  - "llm wiki lint"
  - "wiki lint"
  - "掃 wiki"
  - "wiki 健檢"
argument-hint: "[path]"
---

# llm-wiki-lint — Karpathy LLM Wiki 三層健檢

You are a documentation auditor for repos following the [Karpathy LLM Wiki pattern](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f). Scan the three layers (`raw/` + `wiki/` + `SCHEMA.md` / `index.md` / `log.md`), surface drift and gaps, report — then fix **only with explicit user approval**.

**CRITICAL — 兩段式契約**:
- **Phase 1(report)純 read-only**。掃描期間任何「順手補一下 frontmatter」「順手更新 index」都是違規。
- **Phase 2(fix)必須等 user 對報告明確點頭**(「修」「fix」「都修掉」)才啟動。User 只說「跑 lint」= 只做 Phase 1。

**跟 memory-lint 差異**:
- `memory-lint` → Claude memory 目錄(feedback / user profile / project,prefix-based)
- `llm-wiki-lint` → LLM Wiki repo(wiki/ 主題頁 + frontmatter + raw source traceability + index 一致性)

---

## Step 1: Detect wiki repo path

依序嘗試,命中第一個就用:

| 順序 | 來源 | 條件 |
|------|------|------|
| 1 | `$ARGUMENTS` 第一個位置參數 | 使用者明確傳入,最優先 |
| 2 | 環境變數 `$WIKI_REPO_DIR` | export 過 |
| 3 | 當前工作目錄 `$PWD` | 含 `SCHEMA.md` + `wiki/` 子目錄 |
| 4 | — | 都找不到 → 停止,告訴 user「偵測不到 wiki repo」,**不要瞎猜** |

決定路徑後,先驗證結構:

```bash
ls "$WIKI_PATH/SCHEMA.md" "$WIKI_PATH/wiki/" 2>/dev/null
ls "$WIKI_PATH/index.md" 2>/dev/null || echo "WARN: index.md 不存在"
ls "$WIKI_PATH/log.md" 2>/dev/null   || echo "WARN: log.md 不存在"
```

- `SCHEMA.md` 或 `wiki/` 任一缺 → **fatal**,停止 lint
- `index.md` / `log.md` 缺 → warning,繼續 lint 並在報告建議新建

### Step 1b: 載入 repo 自有規約(lint 的 source of truth)

**不要硬編碼 Karpathy 預設值 — 以目標 repo 自己的宣告為準**:

| 規約 | 來源 | fallback |
|------|------|----------|
| type allow-list | `SCHEMA.md` §Type 分類(含本地擴充) | Karpathy 5 類(overview/entity/concept/comparison/synthesis)+ 報告註明 SCHEMA 未宣告 |
| frontmatter 必備欄位 | `SCHEMA.md` §Frontmatter 欄位 | title / type / last_updated / sources |
| lint 後是否 append log.md | `SCHEMA.md` §Lint 節奏 | 不 append |
| 額外文件風格規則(如「body 禁 date marker」) | repo 的 `CLAUDE.md` / `docs/` 規則檔(如有) | 跳過此項檢查 |

---

## Step 2: Scan & lint(Phase 1 — read-only)

### Step 2a: 掃描範圍

```bash
find "$WIKI_PATH/wiki" -maxdepth 1 -name '*.md' -type f
find "$WIKI_PATH/wiki" -mindepth 1 -maxdepth 1 -type d   # 子目錄偵測 — 上面那條只看 .md,抓不到目錄
```

**包含**:
- `<path>/wiki/*.md` — 所有主題頁(第二條指令找到子目錄時列出並在報告標註,等 user 決定是否納入)
- `<path>/SCHEMA.md` / `index.md` / `log.md` — schema 層
- `<path>/raw/**` — 存在性檢查(gitignored,本地才看得到)

**跳過**:`<path>/journal/`、`<path>/drafts/`、`<path>/scripts/`(非 wiki 一部分)

**工具紀律**:解析 frontmatter / 比對清單一律用 `python3` stdlib(`re` + `glob`),**不要依賴 `yq` / `jq`**(目標機器不保證裝了)。

### Step 2b: Severity decision

| 嚴重度 | 含義 | 範例 |
|--------|------|------|
| 🔴 Error | 結構壞了或宣告與實況矛盾 | 缺必備 frontmatter 欄位、index 列出檔案不存在、body 連結指向已刪頁、SCHEMA.md 數字錯 |
| 🟡 Warning | 沒壞但有腐臭味,user 應該 review | last_updated 漂移、orphan page、stale claim、index 摘要與頁面內容矛盾 |
| 🔵 Info | 啟發式提示,可選處理 | 缺 cross-ref、可能的 missing topic |

不確定時就降一級(warning → info)。**禁止把推測寫成 error**。

### Step 2c: Frontmatter 完整性

每個 `wiki/*.md` 開頭應有 YAML frontmatter,必備欄位以 Step 1b 載入的清單為準:

| 欄位 | 規則 | 缺失等級 |
|------|------|----------|
| `title` | 主題名(string) | 🔴 Error |
| `type` | **在 Step 1b 載入的 allow-list 內** | 🔴 Error(缺欄位 / 不在 allow-list) |
| `last_updated` | YYYY-MM-DD | 🔴 Error(缺)/ 🟡 Warning(格式錯) |
| `sources` | list,指向 `raw/` 或外部來源 | 🔴 Error(缺欄位)/ 🟡 Warning(空 list — 見 Step 2i) |

```bash
python3 - <<'EOF'
import re, glob
for f in sorted(glob.glob("wiki/*.md")):
    s = open(f, encoding="utf-8-sig").read()   # utf-8-sig 吃掉 BOM;\r?\n 容忍 CRLF(Windows 編輯過的檔)
    m = re.match(r"^---\r?\n(.*?)\r?\n---\r?\n", s, re.S)
    if not m: print(f"NO-FRONTMATTER {f}"); continue
    fm = m.group(1)
    for field in ("title", "type", "last_updated", "sources"):
        if not re.search(rf"^{field}:", fm, re.M): print(f"MISSING {field} {f}")
EOF
```

### Step 2d: Index 一致性

| 情況 | 等級 |
|------|------|
| `index.md` 列出但 `wiki/` 下檔案不存在 | 🔴 Error(missing file) |
| `wiki/` 下檔案存在但 `index.md` 沒列 | 🔴 Error(orphan file) |
| `index.md` 的 `last_updated` / `type` 跟檔案 frontmatter 不一致 | 🟡 Warning(drift) |
| `index.md` 標頭宣稱頁數 ≠ 實際頁數 | 🔴 Error |

### Step 2e: Index 摘要 vs 頁面內容矛盾

index 每列通常帶「一句話」摘要。**逐列對照該頁當前核心結論** — 摘要還在講已被頁面推翻的舊事實(例:摘要說「服務 X 運行中、不能停」,頁面已改寫成「服務 X 已停用」)→ 🟡 Warning。這是日期 drift 抓不到的:日期可能同步了、摘要忘了改。

### Step 2f: Body 連結完整性(in-body link integrity)

掃每頁 body 的兩種站內連結,目標不存在 → 🔴 Error:

- markdown link:`[...](some_page.md)` → `wiki/some_page.md` 必須存在(含連字號 / 大寫 / `./` 前綴 / `#anchor` 變體)
- wikilink:`[[some_page]]` 與別名形式 `[[some_page|顯示文字]]` → `wiki/some_page.md` 必須存在

**捕捉要寬不要窄** — 斷鏈檢查的失敗模式是「沉默通過」:regex 漏掉的連結形式不會報錯,user 會誤以為全綠。寬鬆捕捉任何 `.md` 目標,再正規化比對:

```bash
python3 - <<'EOF'
import re, glob, os
for f in sorted(glob.glob("wiki/*.md")):
    s = open(f, encoding="utf-8-sig").read()
    for t in re.findall(r"\]\(([^)#\s]+\.md)(?:#[^)]*)?\)", s):
        if t.startswith(("http://", "https://")): continue   # 外部連結不在此檢查
        rel = t[2:] if t.startswith("./") else t
        if not os.path.exists(os.path.join("wiki", rel)): print(f"BROKEN-LINK {f} -> {t}")
    for t in re.findall(r"\[\[([^\]|#]+)(?:#[^\]|]*)?(?:\|[^\]]*)?\]\]", s):
        if not os.path.exists(f"wiki/{t.strip()}.md"): print(f"BROKEN-WIKILINK {f} -> {t}")
EOF
```

### Step 2g: Stale claims(陳舊聲明)

| 條件 | 等級 |
|------|------|
| git 最後 commit 日 vs frontmatter `last_updated` 差 > 14 天 | 🟡 Warning(frontmatter 沒跟上實際改動) |
| frontmatter `last_updated` > 60 天 且 `type` ∈ {overview, synthesis} | 🟡 Warning(建議 review) |
| 內文含未完成標記 + `last_updated` > 30 天 | 🟡 Warning(可能已有進展沒回寫) |

未完成標記的關鍵詞清單**跟 wiki 語言對齊**(Rule 9):中文 wiki 掃「待 XXX 驗證」「規劃中」「尚未」;英文 wiki 掃 "TBD" / "pending" / "not yet" / "WIP" / "to be verified"。

```bash
# git repo 用 commit 日(mtime 在 clone / checkout 後不可靠);非 git repo 才 fallback stat
git -C "$WIKI_PATH" log -1 --format=%cs -- "wiki/<page>.md"
```

### Step 2h: Orphan pages(孤立頁面)

- 一頁**沒有任何其他主題頁連到它**(只有 `index.md` 列出)→ 🟡 Warning
- `index.md` **不算**入站 ref — index 列出是 Step 2d 的強制要求,每頁都有;若把它當例外,凡通過 2d 的頁面永遠不會是 orphan,本檢查形同虛設

### Step 2i: Data gaps(資料缺口)

| 條件 | 等級 |
|------|------|
| `sources` 欄位空 / 指向消失的 raw 檔案 | 🔴 Error(traceability 斷鏈;raw 可能被搬移過 — 報告附 `find raw -iname` 找到的新位置候選) |
| 主題頁內文 `raw/xxx` 路徑 ref → `raw/**` 實際不存在 | 🔴 Error(local 才能查) |
| 主題頁 Open questions 段 > 3 個月未解 | 🟡 Warning |

### Step 2j: Cross-page contradictions(矛盾)

啟發式(粗略,需人工確認)— **永遠 Warning,不要升 Error**:

- 同一 entity(同一台主機 / 同一服務名)在不同頁描述狀態不一致
- 同一數字(容量 / 並發數 / 頁數)在不同頁不同
- Output 必須註明「需人工確認,可能是更新時差」

### Step 2k: SCHEMA.md 宣告 vs 實況

| 情況 | 等級 |
|------|------|
| SCHEMA.md 說「N 主題頁」但實際 M(N≠M) | 🔴 Error |
| 實際使用的 type 不在 SCHEMA 宣告清單(反向也算:宣告了沒人用不報) | 🟡 Warning — 修法二選一:頁面改 type,或 SCHEMA 擴充收編(報告兩案並列,user 拍板) |

### Step 2l: Repo 自有風格規則(Step 1b 有載入才跑)

例:repo 規定「wiki body 不留 date marker(更新/校正註記),事件日期除外」→ 掃 `(YYYY-MM-DD 驗證)` `(校正)` 類 pattern → 🟡 Warning。**事件本身的日期(「X 月 X 日起」「YYYY-MM-DD 到期」)不算違規** — 逐筆判斷,不確定就標 Info。

### Step 2m: Missing topics / cross-refs(啟發式,永遠 Info)

- `index.md` 某類別空但其他類別有 → 🔵 Info
- 頁 A 內文高頻提到頁 B 的主題詞但沒連結 → 🔵 Info

---

## Step 3: 輸出報告(Phase 1 結束點)

繁體中文,固定格式(英文 wiki 才用英文):

```markdown
# 🔍 LLM Wiki Lint 報告

**掃描目錄:** /path/to/team-wiki
**掃描時間:** YYYY-MM-DD HH:MM
**掃描檔案:** N 個 wiki 頁 + schema 層

## 🔴 Error(建議立即處理)

- [類別] 描述 — 建議動作

## 🟡 Warning(建議 review)

- [類別] 描述

## 🔵 Info(提示,可選處理)

- [類別] 描述

## 🟢 OK(通過)

- Frontmatter 完整性
- ...

## 📊 統計

| Type | 頁數 |
|------|------|
| (依 Step 1b 載入的 allow-list 動態列出實際使用的 type) | N |

**總計:** XX 個 wiki 頁
```

每個 finding 必須包含:

- `[類別]` — Frontmatter / Index / Link / Stale / Orphan / SCHEMA / Contradiction / Style / Data gap
- 具體檔名 + 行為描述
- 建議動作(Error 必有,Warning 可有)

報告結尾固定問一句:**「要修嗎?修的話哪些(全修 / 只修 Error / 指定項)?」** — 不要替 user 決定。

---

## Step 4: Fix(Phase 2 — 必須 user 明確點頭才跑)

User 對報告回「修」後:

### Step 4a: 修機械項

Error + 指定的 Warning,典型:index reconcile(補列 / 砍幽靈列 / 日期同步 / 頁數標頭)、補 frontmatter(從 index 值 + 頁內資料來源段重建)、移除斷鏈、修 sources 路徑、SCHEMA 數字 / type 清單同步。

**修不動的(需 user 判斷的內容矛盾、大範圍風格清理)→ 列為 deferred,寫進 LINT entry,不要硬修。**

### Step 4b: 重跑驗證

Phase 1 同套檢查重跑,機械項應歸零。報告殘留項。

### Step 4c: log.md LINT entry

**僅當 repo SCHEMA 規定 lint 須留痕**(Step 1b 載入)才 append,格式跟著該 repo log.md 既有 entry 走,內容含:findings 分類 + 修了什麼 + deferred 什麼。SCHEMA 沒規定 → 問 user 要不要記。

### Step 4d: Commit

跟著目標 repo 的 commit 慣例(有 PR 流程走 PR)。Commit message 描述 lint 範圍 + 修正類別,不逐項展開。

---

## 範例輸出(虛構 repo,僅示意格式)

```markdown
# 🔍 LLM Wiki Lint 報告

**掃描目錄:** /path/to/team-wiki
**掃描時間:** 2026-01-15 14:20
**掃描檔案:** 17 個 wiki 頁 + SCHEMA.md + index.md + log.md

## 🔴 Error

- [Frontmatter] `wiki/cheatsheet.md` 缺 `sources` 欄位 → 補上來源連結
- [Index] `index.md` 列 `wiki/old_topic.md` 但檔案不存在 → 移除 index 對應行
- [Link] `wiki/onboarding.md` 連到 `data_model.md`(頁已刪)→ 移除或改指現行頁
- [SCHEMA] SCHEMA.md 宣告「15 主題頁」但實際 17 → 更新 SCHEMA.md

## 🟡 Warning

- [Stale] `wiki/load_test.md` last_updated 2026-01-04,但 git 最後 commit 2026-01-13 → 更新 frontmatter
- [Index 摘要] `wiki/service_inventory.md` 摘要寫「服務 X 運行中」,頁面現況已改「服務 X 已停用」→ 改寫摘要
- [Contradiction] `db-host-01` 在 topology.md 寫「gateway + DB」,在 architecture.md 寫「僅 DB」→ 同步描述(需人工確認,可能是更新時差)

## 🔵 Info

- [Cross-ref] `wiki/security.md` 提到「訊息佇列」9 次但沒連 `mq_pipeline.md` → 考慮補連結

## 🟢 OK

- Body 連結完整性:除 1 項均 OK
- Type 分類:全部在 SCHEMA allow-list 內

## 📊 統計

| Type | 頁數 |
|------|------|
| overview | 6 |
| concept | 4 |
| synthesis | 5 |
| sop | 2 |

**總計:** 17 個 wiki 頁

要修嗎?修的話哪些(全修 / 只修 Error / 指定項)?
```

---

## Anti-patterns

- ❌ Phase 1 改任何檔案(連「補個 frontmatter」都不行)
- ❌ User 沒點頭就進 Phase 2(「跑 lint」≠「修」)
- ❌ 硬編碼 Karpathy 5 類當 type 鐵律 — repo SCHEMA 擴充的 type 是合法的
- ❌ 依賴 `yq` / `jq`(用 python3 stdlib)
- ❌ 用 mtime 判 stale(git repo 一律用 commit 日)
- ❌ 訪問外部 API / 抓 Git remote(純 local read;`git log` 是 local)
- ❌ 掃 `journal/` / `drafts/` / `scripts/`(非 wiki 一部分)
- ❌ 路徑偵測不到時瞎猜(必須停止並告訴 user)
- ❌ 把啟發式(cross-ref 缺失、可能矛盾)標成 Error
- ❌ 範例 / 報告裡塞任何真實組織的內部名稱(host / 產品 / 路徑)— 示意一律虛構

---

## Important rules

1. **兩段式契約是核心** — Phase 1 read-only;Phase 2 等明確點頭。違規即破壞 skill 契約
2. **Repo 規約 > skill 預設** — type 清單、必備欄位、log 留痕政策都先讀目標 repo 的 SCHEMA / 規則檔
3. **Path 偵測順序固定** — 不要倒過來、不要跳級、找不到就停
4. **Severity 寧降勿升** — 啟發式 → Info;推測 → Warning;只有結構壞了才 Error
5. **每個 finding 必須有具體檔名** — 不准「某些頁面 frontmatter 不全」這種模糊敘述
6. **Contradiction 永遠標需人工確認** — 你看到的可能只是更新時差
7. **`raw/` 是 gitignored** — 遠端跑(CI / cloud agent)就 skip raw 存在性檢查,並在報告註明
8. **修不動的列 deferred 不硬修** — 內容矛盾、大範圍風格清理要 user 判斷
9. **語言跟 wiki 對齊** — 中文 wiki 用繁中報告,user 明確要求才改
10. **報告 / 範例永遠虛構化** — 不洩任何真實內部 context

More from KerberosClaw/kc_ai_skills

SkillDescription
character-loraUse when the user wants to build a consistent-identity LoRA for an original character — defining the character, generating a face/body-consistent multi-angle dataset (via the gpt-image-gen skill for codex image generation), captioning it, doing base-specific homework, training on a chosen base (Pony / Z-Image / others) on a local GPU, and producing a usable LoRA. This skill ORCHESTRATES the end-to-end pipeline and gates every expensive/irreversible step; it delegates actual image generation to gpt-image-gen and never improvises training settings from memory.
conference-reportUse when the user attended one or more sessions of a conference (with audio recordings + slide photos) and needs help building (1) faithful per-session reconstructions in markdown (slide visuals + speaker transcript with Whisper hallucination annotations), and (2) a downstream report deliverable whose scope and format are decided interactively with the user. Pipeline phase (raw → mlx_whisper Chinese SRT → per-slide multimodal reconstruction → official agenda cross-check via Playwright if available) is deterministic. Report phase is interactive — always quiz user on scope (single-session / single-day / multi-day synthesis), format (existing template / free-form fallback), recipient (formal / informal), and any business workstream mapping before drafting.
ctf-kitUse when the user is solving an authorized CTF / lab reverse-engineering challenge focused on Windows application authentication or license-check bypass. Guides triage → static analysis → dynamic experiment planning → bypass verification, with strong evidence discipline and VM safety boundaries. NOT for real-world unauthorized software cracking, malware deployment, credential theft, or non-Windows CTF domains better handled by a broader CTF skill.
goal-engineerUse when the user wants to AUTHOR an unattended dispatch for either: (1) a goal-driven evaluator-optimizer loop of the GENERATE-AND-SELECT kind (generate candidates → grade against a rubric → iterate by reason-code → keep the best; the human picks the final selection), or (2) a lean build-to-spec run whose build spec is ALREADY FROZEN (approved ADR / locked design / accepted machine-checkable AC) and the only missing piece is the unattended-execution wrapper. A fresh-session agent runs the dispatch hands-off while the human only watches traffic-light push notifications. Interview-style forcing questions lock the spec, then it emits a self-contained dispatch markdown + a channel-agnostic notification protocol. This is the upstream SPEC AUTHOR, NOT a runtime: the dispatch is run by Claude Code's /goal, a headless `-p` session, or any unattended agent — /goal is the engine, this writes what you feed it. It NEVER authors build specs, product decisions, or AC — creating a build spec / PRD from raw input is prd-create. NOT a time scheduler (/loop or cron), NOT a recurring-push registrar (skill-cron).
gpt-image-genUse when the user asks to generate an image via GPT/Codex (e.g. 「叫 gpt 生圖」「幫我用 gpt 生圖」「gpt 畫一個 X」). The skill drafts a Chinese + English prompt pair, iterates with the user until they explicitly approve, then dispatches Codex CLI ($imagegen skill, codex built-in image_gen) in the background, monitors progress, converts the result to a jpg in the current working directory, and writes a sidecar prompt log. Does text-to-image AND img2img — drop a reference image (on-disk file) and it runs Codex `-i` to lock a face/character across scenes.
grillUse when the user says 先討論 / wants to align on a fuzzy idea before building, or before any non-trivial implementation whose requirements are still ambiguous — interviews the user until shared understanding is confirmed, one question at a time, each with a suggested answer; facts answerable from the filesystem/code are looked up, never asked. Optionally maintains the repo's CONTEXT.md glossary when vocabulary settles during discussion. Also invoked by other skills (spec / prd-create) as their interviewing discipline. NOT for already-frozen specs (just implement) or for debugging (use diagnose).
job-scoutUse when the user is considering applying to, interviewing with, or accepting an offer from a company and wants due diligence on the company and optionally a specific role. Requires a company name, optionally a job title, then researches current public data, employee reviews, salary signals, product/tech quality, and red flags before giving a grounded recommendation. NOT for generic career coaching or resume editing.
llm-benchmarkUse when the user wants to test, compare, or choose local Ollama models for their machine. Checks Ollama/GPU state, recommends model sizes from available VRAM, preserves existing benchmark records, pulls only approved models, runs repeatable benchmarks, restores stopped services, and writes a markdown comparison report. NOT for hosted API model evaluation or subjective chat-quality judging without local benchmark commands.
md2pdfUse when the user wants to convert one Markdown file into a publication-ready A4 PDF, especially when the source may contain Mermaid diagrams, ASCII diagrams, CJK text, tables, or pandoc/weasyprint edge cases. Works by copying the source to a _pdf.md working file, converting diagrams, escaping PDF-breaking syntax, rendering with pandoc + weasyprint, then self-checking pages. NOT for batch conversion, slide decks, or editing the original Markdown in place.
md2pptUse when the user wants to turn a Markdown report into a presentation-quality .pptx via interactive design decisions and a reusable hand-coded build script. Drives pre-analysis, global style choices, optional per-slide layout dialogue, python-pptx composition, and optional LibreOffice render self-check. NOT a generic auto-converter, NOT for PDF output, and NOT a fixed brand-template pipeline — brand integration is handled ad hoc through helper primitives.