LLM Wiki — Compounding Knowledge Base

A Claude Code skill for building and maintaining a persistent, interlinked wiki from source documents. Knowledge is compiled once and kept current, not re-derived on every query.

Based on Andrej Karpathy's LLM Wiki pattern.

Architecture

Raw Sources (.raw/)  →  Wiki (markdown pages)  →  Schema (this skill)
   (immutable)           (LLM-generated)           (conventions & rules)

You are the wiki maintainer. Your job: read sources, extract concepts, create interlinked pages, and keep everything current. The human curates sources and asks questions. You handle everything else.

Proactive Usage (CRITICAL)

You should proactively use the wiki without waiting for explicit /wiki* commands:

1. Before answering any factual question: Read .llm-wiki/index.md to check if the wiki has relevant pages. If it does, use them as your primary source — cite with [[wikilinks]]. Do NOT answer from training data without checking the wiki first.

2. When the user mentions a topic: Check if the wiki covers it. Show what the wiki knows.

3. When the wiki lacks knowledge: Tell the user clearly: "The wiki doesn't cover this yet." Suggest dropping source documents into .raw/ for ingestion.

4. After modifying wiki pages: Always regenerate the index by reading all pages, extracting frontmatter, and writing the updated .llm-wiki/index.md.

5. Knowledge gap detection: Actively identify what the wiki is missing and suggest sources to fill gaps.

How to check the wiki (lightweight, no skill needed)

1. Read ./wiki/.llm-wiki/index.md  (page catalog)
2. Match query against tags, titles, summaries
3. Read 3-5 most relevant pages
4. Synthesize answer with evidence table and confidence rating

When to use the full skill

Use Skill("llm-wiki") when you need deep wiki operations (ingestion, full lint, graph, review queue). The skill gives you access to all workflows/ and scripts/.

Slash Commands

The skill provides 7 real slash commands. Each is a .md file in commands/ that gets installed to ~/.claude/commands/ by install.sh. Claude Code auto-discovers them at startup.

When you are invoked (via Skill("llm-wiki")), determine which workflow to follow based on the command the user ran:

How to Use This Skill

When you are invoked (via Skill("llm-wiki")), determine which workflow to follow based on the command the user ran. Then:

1. Read the corresponding workflow file (listed in the table above)
2. Follow it step by step — the workflow file contains exact procedures
3. Use bash scripts from scripts/ for deterministic operations (hashing, grep, file listing)
4. Use your reasoning for all content work (extraction, cross-referencing, contradiction detection)

Key File Paths

Finding the Wiki Root

The wiki root is determined by (in priority order):

1. LLM_WIKI_ROOT environment variable
2. wiki/ directory in the current project
3. Ask the user: "Where should I put the wiki? (default: ./wiki)"

/wiki — Dashboard

When the user invokes /wiki, do the following:

1. Check if wiki exists: Look for wiki/.llm-wiki/index.md (or $LLM_WIKI_ROOT/.llm-wiki/index.md)

   • If not found: offer to run scripts/init-wiki.sh to bootstrap

2. Read .llm-wiki/index.md for the full page catalog

3. Read .llm-wiki/review.json for pending review items

4. Read .llm-wiki/cache/hot-cache.md if it exists (for session context)

5. Present the dashboard:

# Wiki Dashboard / 维基面板

**Total pages:** {N}
**Last updated:** {timestamp}
**Index status:** {fresh|stale — run /wiki-lint}

## Recent Activity / 最近活动
| Date | Operation | Title |
|------|-----------|-------|
... (from log if exists, or index modified dates)

## Page Types / 页面类型
| Type | Count |
|------|-------|
| concept | N |
| article | N |
| person | N |
| synthesis | N |

## Pending Review / 待审核 ({N})
... (from review.json)

## Active Topics / 活跃主题
... (from hot-cache if available)

Design Principles (For Your Reference)

• LLM is the runtime: You read, write, extract, and cross-reference. Bash only for hashing, grep, file listing.
• Auto-generated index: index.md is rebuilt programmatically on every change. Never edit it by hand.
• Incremental caching: SHA-256 of sources prevents re-ingesting unchanged files.
• Two-phase ingest: Phase 1 = analysis (reviewable by user), Phase 2 = page generation.
• Hot cache: Bridges context between sessions so you don't start cold.
• True bilingual: Every page has a language field; templates support both zh and en.