AutoGineer
Generated banner for @micuintus/llm-wiki
Generated logo for @micuintus/llm-wiki

@micuintus/llm-wiki

LLM wiki kit

First committed by micuintus

Published in micuintus/llm-wiki. Publisher page

32 GitHub stars

What it is

As minimal as Pi. Karpathy's LLM Wiki pattern as a minimal, agent-agnostic skill.
— the description by micuintus

AutoGineer overview, from the brain’s own files

llm-wiki is a lightweight, dependency-free skill designed to implement the Karpathy LLM Wiki pattern. It functions as a persistent memory layer where an AI agent curates, summarizes, and cross-links information from raw sources into a structured knowledge base. By relying on conventions defined within the skill text rather than complex code or external infrastructure, it ensures that data remains portable and readable by any agent. The system consists of a single core skill file and a set of lazy-loaded reference files, keeping the footprint minimal while allowing for complex knowledge compounding.

How it works

Memory model: LLM wiki

AutoGineer overview, from the brain’s own files

The system operates by placing a single SKILL.md file into an agent's skill directory. This file contains the instructions and conventions the agent follows to process information. It uses a references folder containing lazy-loaded recipes for specific tasks, such as handling session transcripts or code repositories. The agent reads these instructions to summarize, link, and maintain the wiki, which is stored as plain markdown files on the user's disk.

Folders
FolderRoleFiles
llm-wikifolder8

Requirements

Runtimes: Node.js

Databases: SQLite

Set it up

  1. Download the brain as a ZIP and unzip it where you keep your notes or project: https://mp-staging.autogineer.com/agent-brains/micuintus/llm-wiki/download

  2. cd llm-wiki && claude

  3. AutoGineer scanned these files at commit e03cd38.

  4. First committed by micuintus. Published in micuintus/llm-wiki. Licence: MIT License.

Usage

AutoGineer overview, from the brain’s own files

  • Curating research papers into a searchable, interlinked knowledge base.
  • Documenting project architecture, decisions, and open questions.
  • Mapping complex ecosystems or taxonomies through agent-assisted ingestion.
  • Synthesizing book chapters into concept pages and theme threads.
“Summarize the last five research papers I uploaded and link them to the existing 'Quantum Computing' concept page.”
The agent updates the 'Quantum Computing' markdown file with new summaries and creates cross-references to the specific paper files.
“Review my project documentation and identify any contradictions between the architecture decisions and the current bug list.”
The agent scans the wiki files and flags conflicting information, suggesting updates to maintain consistency.

Documentation by micuintus

@micuintus/llm-wiki

As minimal as Pi.

Karpathy's LLM Wiki pattern as a skill. One file. ~6 KB. No dependencies, no CLI, no web server, no Obsidian plugin.

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

Install

Pi (npm):

pi install npm:@micuintus/llm-wiki

Pi (GitHub):

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

Everyone else:

# Copy the skill directory (SKILL.md + its lazy-loaded recipes). Start ingesting.
curl -L https://github.com/micuintus/llm-wiki/archive/refs/heads/main.tar.gz \
  | tar -xz --strip-components=1 -C ~/your-agent/skills/ llm-wiki-main/llm-wiki

That lands ~/your-agent/skills/llm-wiki/SKILL.md with its references/ alongside, so the lazy-loaded recipes resolve. (SKILL.md is the only always-loaded file — the rest is read on demand.)

What you get

One file: SKILL.md. Everything else is lazy-loaded — referenced by path in the skill text, never loaded until you actually need it.

Agent session recipes (lazy-loaded):

  • Pi sessions — JSONL trees with fork detection

  • Claude Code / opencode / Gemini CLI — transcripts, SQLite, JSON

Code-source recipe (lazy-loaded) — grep+cite-by-line conventions for when the source is a repository rather than prose.

Web chat ingestion — Claude.ai, ChatGPT, Gemini via CDP. A separate package: ingest-web-chat (pi install npm:@micuintus/ingest-web-chat).

The lazy-loaded subskills sleep forever until you wake them. (Don't forget to kiss these sleeping beauties.)

Use cases

  • Reading a book — chapter summaries, character/concept pages, theme threads, all interlinked.

  • Researching a topic across many papers — concept pages cite multiple sources; contradictions surface instead of getting lost.

  • Mapping an ecosystem (e.g., the Pi extension landscape) — taxonomies, comparisons, primitive mappings.

  • Documenting your own project — decisions, architecture, bugs, open questions, all linked through one schema.

Why this exists

Most wiki implementations ship a CLI, a web server, a static site generator, and 2,000 tests. That's fine if you want infrastructure. This is for people who want a pattern — a convention the agent already understands — and nothing else.

Conventions live in the skill text, not in code. The wiki itself is plain markdown on disk — if you ever switch agents or skills, your data travels with you unchanged.

Related work

The Karpathy gist has spawned a diverse ecosystem. Every implementation below shares the same three-layer foundation (raw sources → compiled wiki → schema); they differ in shape (how the agent consumes the pattern) and weight (how much tooling ships alongside the skill).

What makes @micuintus/llm-wiki different: it is a pure skill — a single SKILL.md (~6 KB) plus lazy-loaded reference files that any agent (Pi, Claude Code, Codex, Cursor) can read directly. No runtime dependencies, no generated code, no CLI to install, no web server to run, no Obsidian plugin to manage. Conventions are enforced by the skill text itself, not by code. If you want the lightest possible commitment to the pattern — copy one file, start ingesting — this is it. Other implementations add value in exchange for weight: IDE plugins, static-site generators, code-based guardrails, or self-driving agents. Pick the shape that matches your workflow.

RepoStarsContributorsLicenseLanguageShapeNotes
Astro-Han/karpathy-llm-wiki6511MITMarkdownSkillAgent Skills-compatible. Single-file SKILL.md, no session recipes. Closest to a vanilla skill.
toolboxmd/karpathy-wiki661—ShellSkillTwo Claude Code skills: setup + maintenance. Hooks-driven.
balukosuri/llm-wiki-karpathy1251——SkillArticle + implementation walkthrough.
lewislulu/llm-wiki-skill4472MITTypeScriptSkill + ecosystemSkill + Obsidian audit plugin + local web viewer + shared TypeScript audit library. For users who want IDE-like tooling.
lucasastorian/llmwiki808—Apache-2.0Python + TypeScriptCLI toolUpload documents, connect Claude via MCP, agent writes the wiki. MCP-native.
Pratiyush/llm-wiki2291MITPythonCLI toolllmwiki CLI: session ingestion, static site generation, 2,651 tests, 16 lint rules, MCP server, Playwright E2E, AI exports. Most feature-rich; heaviest.
Kausik-A/pi-llm-wiki91MITTypeScriptPi-nativePi package with bundled skill + Pi extension. Enforces guardrails via code rather than convention.
yologdev/karpathy-llm-wiki432—TypeScriptSelf-growing"Yoyo" AI agent grows the wiki from Karpathy's founding prompt. Commits are the agent's work. Includes web viewer.
hsuanguo/llm-wiki112MITPythonSelf-growingWiki that "evolves with you" + Python scaffolding.

Changelog

See CHANGELOG.md.

License

MIT

CHANGELOG.md

Changelog

[1.4.12] — 2026-10-26

Changed

  • Package author and LICENSE copyright now read Michael Voigt (micu), matching the other packages.

[1.4.11] — 2026-10-25

Removed

  • ingest-web-chat stub. A second skill with the same name collides with the real one once both are installed. Install @micuintus/ingest-web-chat for web chat ingestion.

[1.4.10] — 2026-10-25

Changed

  • ingest-web-chat moves out into its own repo and package (micuintus/ingest-web-chat, npm @micuintus/ingest-web-chat), extracted with full history. Install it with pi install npm:@micuintus/ingest-web-chat. The old path keeps a stub pointer for this release only.

[1.4.9] — 2026-10-23

Added

  • ChatGPT extractor for ingest-web-chat: dedicated provider alongside the Claude/Gemini fallback, with a gate guard (a Cloudflare challenge page was once saved as a chat and reported success) and CDP auto-launch for cold Chrome starts.

Fixed

  • Subskill discovery: Pi treats a directory containing SKILL.md as a skill root and does not recurse, so llm-wiki/skills/ingest-web-chat was never discovered; folded the shadowing shim in and exposed it in the manifest.

  • Scroll Gemini history to top before extracting.

[1.4.8] — 2026-05-30

Fixed

  • Non-Pi install now fetches the whole llm-wiki/ directory (tarball → tar) instead of just SKILL.md. The single-file copy left the lazy-loaded references/* recipes unresolvable; the skill pointed at paths that weren't there.

[1.4.7] — 2026-05-30

Fixed

  • Page-type consistency — pi-session-recipe.md used type: session, which isn't in the canonical type list and would be rejected by the type-validity lint; switched to source (matching the agent-session recipe). SCHEMA.template.md page-types list was missing stub.

Added

  • log.md vs git clarified in SKILL.md: it's the VCS-independent, semantic operation log — keep entries about what knowledge changed, not the file diffs git already records.

  • Parallel ingest note: dispatch one subagent per independent source, but serialize writes to shared files (index.md, log.md, raw-sources/index.md).

[1.4.6] — 2026-05-30

Added

  • Confidentiality guidance in SKILL.md: for proprietary or confidential sources, prefer reference-by-path over copying and don't reproduce confidential identifiers in any page you may publish.

Changed

  • references/code-source-recipe.md examples now use a fictional generic engine rather than identifiers from a private codebase.

[1.4.3] — 2026-05-09

Fixed

  • README install command — pi install github:micuintus/llm-wiki doesn't work; pi has no github: source handler (only npm:, git:, and protocol URLs). Replaced with pi install git:github.com/micuintus/llm-wiki.

  • references/pi-session-recipe.md — verification preamble now references @earendil-works/pi-coding-agent (the package was renamed from @mariozechner/pi-coding-agent; session layout is unchanged).

[1.4.2] — 2026-05-07

Fixed

  • str_replace → "targeted edits" in SKILL.md Compile step. Pi's tool is named edit; the term str_replace is Anthropic-tool-API jargon. When the model reads str_replace in the skill text, it may look for a tool that doesn't exist in Pi. Reworded to be tool-agnostic. (Thanks to a friendly reviewer for spotting.)

[1.4.1] — 2026-05-07

Added

  • Tagline: "As minimal as Pi." Added to README subtitle and package.json description.

  • Pi GitHub install: pi install github:micuintus/llm-wiki alongside existing npm install.

Changed

  • README — Restructured Install section with three explicit paths: Pi (npm), Pi (GitHub), everyone else (curl).

  • package.json keywords — Added "minimal".

[1.4.0] — 2026-05-07

Added

  • references/code-source-recipe.md — lazy-loaded recipe for ingesting code repositories: register-by-path (don't copy), grep-driven inventory, path:line citations, source-size heuristic (~1 page per ~1000 lines per cohesive subsystem), commit-SHA recording. Closes a gap exposed by real-world stress testing on a 6,400-line C++ engine source.

  • references/synthesis-and-archive.md — lazy-loaded recipe for filing query answers: synthesis vs. archive distinction, when each cascade-updates, decision rule ("if value disappears with citation changes, it's synthesis; else archive").

  • PDF guidance in SKILL.md Special sources: extract to text first, register the extracted text path.

  • Source-size heuristic in quality.md: ~1 page per ~1000 lines is a useful starting point; under-mining is the common failure.

Changed

  • SKILL.md trimmed 6,957 → 6,260 bytes by lazy-loading the archive-on-request workflow and the correction pattern. Both are now one-line pointers in the Query section.

  • README byte claim corrected: "~4 KB" → "~6 KB". The previous figure was stale since 1.2.x growth.

  • Special sources section in SKILL.md reorganized: code repositories first (most common case), then sessions, web-chat, binaries, PDFs. Reliability paragraph moved into the same section.

Why

Review after a 49-page real-world code ingest found three gaps: code-as-source had no recipe, the README's byte claim was stale, and two SKILL.md subsections (archive workflow, correction pattern) were content that earned their place but didn't earn always-loaded bytes.

[1.3.0] — 2026-05-07

Added

  • type: stub for registered-but-not-compiled sources and forward-reference placeholders. Lint flags stubs older than 30 days. Depth rule: frontmatter + one-paragraph rationale + link to raw source.

  • Source reliability protocol. New reliability: high|mixed|unverified field for sources, with documented per-claim attribution pattern in quality.md for compiling from mixed/unverified material (typical of LLM chat exports). Lint flags mixed/unverified sources cited in compiled pages without a ## Source reliability section.

  • Archive-on-request workflow in Query section. Explicit rule: synthesised answers always become new pages (never merged into existing concept pages); index summary prefixed [Synthesis] or [Archive]; logged as archive verb.

Changed

  • Lint rewritten with deterministic / heuristic split. Auto-fix list (index-FS sync, internal links, raw refs, See Also bidirectionality, frontmatter validity, stub aging) separated from report-only list (contradictions, stale claims, orphans, concept gaps, unattributed mixed-source claims, archive-page drift). Sharper authority boundary; modeled after Astro-Han/karpathy-llm-wiki's lint structure.

Fixed

  • ingest-web-chat: ## Conversations casing bug. Section header detection was case-sensitive and would write a duplicate capital-C section if the existing wiki used lowercase. Now case-insensitive match preserves whatever casing the wiki already uses.

[1.2.3] — 2026-05-07

Changed

  • Page types collapsed back to Karpathy's 7. Removed taxonomy and implementation from the canonical type list; they're now concept variants documented in quality.md with their own depth rules. Same depth bar, less type proliferation.

  • Sharpened trigger phrases: "add to wiki", "save this", "file this", "remember this", "ingest this paper/article/chat/video", "what do I know about X", "knowledge base". Broader and more natural than the old "ingest into the wiki" / "lint the wiki" set.

  • Session recipes are self-correcting: added Verification preamble to pi-session-recipe.md and agent-session-recipe.md instructing the LLM to peek at one session file and confirm the schema before applying. If drifted, adapt the recipe and update the preamble — don't silently work around it.

Removed

  • Cross-wiki links section from SKILL.md. The lint rule "every ](path.md) must resolve from the file's directory" already covers it; the worked example was just path arithmetic.

Fixed

  • Tarball bloat: nested .npmignore excludes package-lock.json from the ingest-web-chat subskill on publish. Tarball: 59.6kB → 41.7kB.

[1.2.0] — 2026-05-07

Added

  • New page types: taxonomy (categorized enumeration with comparison matrix) and implementation (concrete code with architecture diagram + design decisions table).

  • Tags frontmatter: Optional but recommended tags: [tag1, tag2] for discoverability on wikis >20 pages.

  • Correction pattern: Document corrections explicitly rather than silently rewriting — preserves learning arc.

  • Cross-wiki links: Convention for sibling-project wikis at same repo level (../../../../sibling/llm-wiki/page.md).

  • Relative path verification lint rule: Every ](*.md) link must resolve from the file's directory.

  • Bidirectional link lint rule: If page A links to page B, page B must link back.

  • Tag presence lint rule: Warn if tags: is missing on pages >100 lines.

  • Proper CHANGELOG.md: Generated from full git history (1.0.0 → 1.2.0).

Changed

  • Depth rules (references/quality.md): Added minimums for taxonomy and implementation.

  • SCHEMA template (references/SCHEMA.template.md): Added taxonomy/implementation types, tags convention, lint rules.

  • SKILL.md: Added new types, tags, correction pattern, cross-wiki links, path verification.

  • README: Rewritten — cut bloat, replaced with direct opinion. Session recipes and web chat listed as lazy-loaded core features.

[1.1.5] — 2026-05-05

Added

  • Ingest vocabulary: register vs compile vs ingest with precise log conventions.

  • Raw-sources format: raw-sources/index.md registry with path | slug | topics table.

  • Forward references: Links to not-yet-compiled pages are explicitly allowed.

  • HTML/auth source handling: Acknowledged in Register step.

  • Related work table: Single table with Shape column and differentiation paragraph.

Changed

  • README: Related work consolidated into one table with Shape and Notes columns.

[1.1.2] — 2026-04-30

Added

  • ingest-web-chat subskill: CDP-driven web chat ingestion from Claude.ai, ChatGPT, Gemini, Le Chat.

Changed

  • SKILL.md slimmed: 252 → 77 lines. Quality heuristics and tooling lazy-loaded from references/.

  • Subskill cleanup: Dropped dead code, unused deps, orphan refs.

[1.0.2] — 2026-04-29

Added

  • Page quality heuristics (references/quality.md): Minimum depth per type so pages don't stay stubs.

  • Query filing: Syntheses that connect ≥2 pages get offered as type: synthesis.

  • 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.

Changed

  • README: Clarified minimal design, lazy-loaded recipes, Karpathy authority.

[1.0.1] — 2026-04-28

Added

  • Agent session recipe (references/agent-session-recipe.md): Claude Code JSONL, Gemini CLI JSON, opencode SQLite.

  • Pi session fork traversal (references/pi-session-recipe.md): Fork detection, custom event extraction, branch-by-branch reading.

  • npm install instructions in README.

Changed

  • SKILL.md: Auto-fix deterministic lint, dropped search paragraph.

[1.0.0] — 2026-04-28

Added

  • Initial release: Karpathy's LLM Wiki pattern as a minimal skill.

  • Three operations: ingest, query, lint.

  • Three-layer architecture: raw sources → compiled wiki → schema.

  • 7 canonical page types: concept, decision, bug, open-question, source, reference, synthesis.

  • Pi package manifest for pi install.

16 files

Advantages

Limitations

  • Lacks built-in IDE plugins or graphical interfaces for those preferring visual tooling. evidence: README.md
  • Relies on agent-driven convention adherence rather than hard-coded guardrails. evidence: README.md

Security

Risk level: No risk found

This brain runs

Agent instructions that run or fetch something (2)

  • llm-wiki/SKILL.md: git already records the diff.
  • llm-wiki/SKILL.md: For many independent sources, ingest can run in parallel: dispatch one subagent per source to compile concurrently.

Attribution and licence

First committed by
GitHub avatar of micuintusmicuintus
Declared authors
  • Michael Voigt (micu) (declared in package.json)
Publisher
micu (Community)
Source
micuintus/llm-wiki · commit e03cd38
Licence
MIT License (files served here)
  • ATTRIBUTION.md (written by AutoGineer)
Synced
Contributors
micuintus

AutoGineer is not affiliated with or endorsed by micuintus.

Download

The ZIP holds the files AutoGineer serves, with ATTRIBUTION.md and the licence. Download llm-wiki.zip