@micuintus/llm-wiki

Karpathy's LLM Wiki pattern as a minimal, razor-sharp skill for Pi and other agents.

The user curates sources. The LLM does the bookkeeping — summarizing, cross-linking, flagging contradictions. Knowledge compounds in the wiki rather than being re-derived from raw chunks on every query.

Design philosophy

Minimal. No dependencies, no JSON metadata, no enforced extensions. Standard markdown, standard links, standard frontmatter. Convention-based guardrails, not code-based ones.

Razor-sharp. Every line in SKILL.md serves the pattern. No bloat, no abstractions that don't pay for themselves.

Karpathy's gist is authoritative. The skill implements his three-layer architecture (raw sources → wiki → schema) and three operations (ingest, query, lint) faithfully. Where we diverge, we diverge minimally and document why.

Agent session ingestion is first-class. Pi JSONL trees (with fork detection), Claude Code JSONL, opencode SQLite, and Gemini CLI JSON are all supported. Session recipes are lazy-loaded — referenced by path in SKILL.md, not inlined. They only load when you actually scan sessions, never on general skill load.

Web LLM chats are first-class. The ingest-web-chat subskill imports single chats from Claude.ai, ChatGPT, Gemini, and Le Chat by URL via CDP against a real Chrome — the only reliable path past Cloudflare Turnstile and enterprise SSO.

Install

Pi (npm — recommended)

pi install npm:@micuintus/llm-wiki

Pi (git)

pi install https://github.com/micuintus/llm-wiki

Claude Code / Codex / other agents

Copy llm-wiki/SKILL.md and llm-wiki/references/ into your agent's skills directory.

What it's for

• Book and paper research — ingest chapters, papers, figures; build interlinked concept pages as you read
• Agent session preservation — Pi JSONL sessions with tree traversal and custom event extraction; Claude Code / opencode / Gemini CLI transcripts with subagent support
• AI model development — code, checkpoints, datasets, audio/MIDI with companion descriptions
• Software porting documentation — track architecture research, design decisions, and porting progress

Structure

llm-wiki/
├── SKILL.md              # skill instructions (~4 KB, minimal)
├── references/
│   ├── source.template.md    # raw source copy template
│   ├── SCHEMA.template.md    # per-project schema skeleton
│   ├── quality.md            # depth rules per page type
│   ├── pi-session-recipe.md  # Pi JSONL fork detection + extraction
│   └── agent-session-recipe.md  # Claude Code, opencode, Gemini CLI
└── skills/
    └── ingest-web-chat/      # CDP-driven web chat ingestion

What's in SKILL.md

• Page types — 7 canonical types (concept, decision, bug, open-question, source, reference, synthesis)
• Page quality heuristics — minimum depth per type so pages don't stay stubs (lazy-loaded from references/quality.md)
• Query filing — syntheses that connect ≥2 pages get offered as type: synthesis so good answers don't disappear into chat history
• Lint rules — deterministic fixes (broken links, orphans, type consistency) + heuristic reports (stale claims, thin pages, concept gaps)
• Schema co-evolution — triggers that prompt SCHEMA.md updates

Session recipes (lazy-loaded)

SKILL.md references two recipe files by path. They are not loaded on general skill load — only when you actually need to scan agent sessions.

• references/pi-session-recipe.md — Pi JSONL tree traversal. Fork detection (mandatory), custom event extraction (fetch failures, rate limits, tool artifacts), branch-by-branch reading.
• references/agent-session-recipe.md — Claude Code JSONL, Gemini CLI JSON, opencode SQLite (dual schema variants). Inventory, triage, extraction patterns. One-line table for quick tool→location mapping.

Related work

Other implementations of Karpathy's LLM Wiki pattern, each with different trade-offs:

• Astro-Han/karpathy-llm-wiki — the most widely used pure-markdown skill. Agent Skills-compatible for Claude Code, Cursor, and Codex. Good starting point if you want a single-file skill without session recipes.
• lewislulu/llm-wiki-skill — skill plus Obsidian plugin, web viewer, and audit plugin. Heavier but richer ecosystem if you live in Obsidian.
• Pratiyush/llm-wiki — a Python CLI tool (llmwiki) that ingests sessions and generates a static site. Stdlib-only, exports llms.txt / llms-full.txt. Different shape (tool vs skill) but same pattern.
• toolboxmd/karpathy-wiki — two Claude Code skills (setup + maintain). Hooks-driven, simple.
• Kausik-A/pi-llm-wiki / iRonin/pi-llm-wiki — Pi-native packages with a bundled skill and a Pi extension for deterministic operations, guardrails, and generated metadata. Convention vs enforcement trade-off: they enforce via code; this skill enforces via convention.
• yologdev/karpathy-llm-wiki — self-growing wiki via AI agent "yoyo". Includes a web viewer.

License

MIT