Karpathy Pattern

Community implementations of Andrej Karpathy’s LLM-maintained wiki pattern — the same pattern this vault is built on. Tracks other people’s open-source vault templates, variant schemas, and structural choices so we can compare approaches and harvest good ideas.

This topic sits alongside Wiki Community Enhancements (a research survey of the broader ecosystem) but focuses on specific, individually-reviewed implementations.

The two endpoints of the spectrum are well-represented: the Stride starter (minimal Obsidian template, 4 operations) and synthadoc (architecturally complete Python engine + Obsidian plugin, ~10 operations, SQLite audit DB). When a third-party tool offers a pattern this vault should adopt, the analysis lives at Karpathy Wiki Additions from Synthadoc.

Articles

• karpathy-obsidian-vault) — Minimal faithful implementation of Karpathy’s pattern. Obsidian template with a 4-operation CLAUDE.md (Ingest, Research, Query, Lint). Ships empty, created as a “hello world” by Stride AI Academy. 9 stars.

• Synthadoc — LLM-Powered Wiki Engine — Architecturally complete Python engine + Obsidian plugin implementation by axoviq-ai. v0.3.0 / 231 stars / AGPL-3.0 / 2026-05-04. 5-pass IngestAgent, status frontmatter, query decomposition, hooks, SQLite audit DB. The opposite-end-of-the-spectrum from the Stride starter — formal engine vs minimal template.

• Karpathy Wiki Additions from Synthadoc (2026-05-05) — The specific 5 improvements this vault adopted from reading synthadoc’s design doc, what was deliberately skipped, and why. Includes status frontmatter field, bin/lint-stale-sources, bin/build-audit-db, query decomposition behavior, and per-topic AGENTS.md.

• Pipeline Diagram — How This Wiki Actually Works (2026-05-20) — Single-image architecture diagram of the full pipeline. Six columns left-to-right (Inputs → Staging → Compile → Wiki → Publish → Live) with the /inbox-refresh skill prominent at the top of the fan-in, the three-way triage: fork (ingest | refresh | skip) inside Compile, a recurring Maintenance band on top (Lint / Refresh / Watch / Cross-link / Connection), QMD as the retrieval substrate underneath (7-stage pipeline broken out), and bin/ scripts as the operational layer. Self-contained ~52KB SVG, currently v3 (2026-05-20); v1 (2026-05-09) + v2 (2026-05-15) preserved. Date-anchored — newer snapshots get fresh dated filenames, old ones preserved.

• Build Karpathy’s LLM Wiki for Your Business — 10-Minute Walkthrough (Dream Labs AI) — Third-party YouTube tutorial (FAWm7DuFSPc) on adopting the Karpathy LLM-wiki pattern for SMB use. Drop Karpathy’s GitHub gist into a fresh Obsidian vault as CLAUDE.md → Claude Code scaffolds wiki structure → drag raw business data (PDFs, transcripts, goals, competitor research) into raw/ → compile in ~9 minutes. Includes the Alex Hormozi 12-question self-interview as a seed-content pattern, the gap-detection query (“Tell me where the gaps in my knowledge are”), and a work-in-progress role-model-mentor skills-plugin concept (“like a Nintendo gaming cartridge”). Validates this vault’s pattern from the outside; documents how the 20-topic schema diverges from Karpathy’s default 4-folder (concepts/entities/sources/synthesis) layout for multi-topic scale.

• Build an AI Second Brain — Wiki + Journal + CRM (Matt Wolfe) — Third-party YouTube build (yke4fLQUsh4) extending the LLM-wiki pattern with two net-new layers: a journal whose AI responses are grounded in the saved wiki + prior entries + CRM, and a CRM keyed on person names. Capture via the Obsidian Web Clipper (auto-pulls YouTube transcripts) into raw/; agent runtime is Codex (Claude Code / Cowork noted as equivalents); agents.md is the whole control surface; a Codex hourly Automation + GitHub push make ingestion and backup run on autopilot. Distinct from the Dream Labs walkthrough (Claude Code + Hormozi seed, wiki-only) — this adds the journal + CRM layers and the automation loop.

• Agent Wikis — Compiled LLM Wikis with Agent-Native Access — Commercial hosted service (agentwikis.com) building and maintaining 15+ public wikis on the Karpathy pattern. Built by a Hermes Agent dev; explicitly credits Karpathy’s gist. The first implementation with published empirical benchmarks: compiled wiki 89% correct / 7% hallucination vs web search 48% / 48% on niche fast-moving domains (27-task blind-judged eval). Free public wikis + llms.txt + MCP server; sponsor a topic for $50 one-time.

• Google Open Knowledge Format (OKF) — the LLM-Wiki Pattern, Standardized — Google Cloud’s open specification (2026-06-12, Sam McVeety + Amir Hormati) representing knowledge as a directory of markdown files with YAML frontmatter; explicitly cites Karpathy and “formalizes the LLM-wiki pattern into a portable, interoperable format.” One required frontmatter field (type); reserved index.md + log.md — the same two load-bearing files this vault runs; no schema registry, no central authority. The first vendor-backed standardization of the pattern (a stronger signal than any community implementation here). Ships a reference enrichment agent (Google ADK + Gemini over BigQuery) + a backend-free HTML visualizer + three browsable bundles. This vault is near-conformant already; the gap is OKF’s type-as-concept-kind vs this vault’s type-as-pipeline-stage.

• Every Level of a Claude Second Brain (Nate Herk) — five-level taxonomy of second-brain retrieval sophistication (CLAUDE.md routing → LLM wiki → semantic/vector search → knowledge graph → always-on OS), with the thesis “pick the lowest level that fixes a real pain point.” A decision-framework counterpart to the build tutorials above.

• GBrain — Garry Tan’s Open-Source AI Brain — open-source personal AI brain: a self-wiring regex knowledge graph over markdown source-of-truth, hybrid retrieval, and a nightly “dream cycle” (+31.4 pts precision@5 over vector RAG in the creator’s evals); YC’s Summer-2026 RFS names the company version (“every company needs a brain”) a fundable category. Two videos (YC/strategy + architecture/evals).

• Fable 5 Second Brain — A Four-Layer Agentic Operating System — a creator has Claude Fable 5 build a visual second brain organized around four structural components (Applications / Routines / Memory / Skills — a what’s-connected inventory, distinct from the five-level retrieval-sophistication taxonomy above). Central claim: ~40% token savings (30K vs 50K, verified via /context) from brain.js, a deterministic keyword-scoring retrieval layer that ranks and opens files without invoking the model until the final answer. Fed Fable three named inspirations — QMD, GBrain, and Graphify — plus /last30days research on current second-brain best practices.

• Pinecone Nexus — A Compiled Knowledge Layer (Karpathy LLM-Wiki Parallel) — Pinecone (the vector-DB vendor) admits agentic RAG has fundamental problems (~85% of agent effort on retrieval, 50–60% task completion, runaway tokens) and ships Nexus, a compiled knowledge engine that moves reasoning from query-time to ingest-time. A third-party walkthrough maps it point-for-point onto Karpathy’s LLM Wiki: compiled “artifacts” ≈ wiki markdown files, a “context compiler” coding agent ≈ the maintainer + CLAUDE.md, a contradiction-auditing “bookkeeper,” declarative NoQL with field-level citations, eval-driven artifact generation (works only for known/repeatable tasks). Google Cloud Knowledge Catalog + Microsoft Fabric IQ are the same trend. The strongest vendor signal yet that compile-up-front is becoming a category — landing on the ingest-time side of the Wiki-vs-Open-Brain fork.

• Karpathy’s Wiki vs. Open Brain — single-creator (Open Brain’s author) critique framing AI memory as an ingest-time vs query-time fork: Karpathy’s wiki compiles/synthesizes at ingest (wins for solo deep research, but can drift into confident “active misinformation” and strains for teams / >10k docs), vs a SQL-backed store that synthesizes at query (provenance + scale); proposes a never-edited wiki view generated from the database. The critique-from-outside of this vault’s own pattern.

• Open Skills — Portable Agent Procedures (Nate B Jones) — the procedure-layer companion to the same creator’s Open Brain. Portable skill.md procedures defined by a six-element contract (trigger / anti-trigger / owned-job / tools / boundaries + output / verification), composed into runbooks, and grown by a session-to-skill extractor flywheel that mines completed sessions into reusable skills. Where the wiki pattern is about what the agent knows, Open Skills is about what the agent does — the how-to layer that sits beside the knowledge layer.

• Open Engine — Cross-Agent Coordination Layer (Nate B Jones) — the third entry in Jones’s “Open ___” franchise (Open Brain = memory → Open Skills = procedures → Open Engine = coordination/handoff). Uses a shared issue tracker (Linear) as a vendor-neutral coordination bus plus four portable skills (setup / status / run-a-queue / smoke-test): an agent claim-locks an “agent instructions”-labeled issue, moves it through status states (to-do → working → needs-input → done), and leaves an “agent claimed” receipt; one teammate’s Codex agent can hand a self-contained ticket to another’s Claude agent with no direct integration. Reframes “ticket vs prompt” (a ticket carries owner + background + allowed actions + stop point + required proof). Caveat: not open-source despite the branding — skills gated behind his Substack/Slack.

• Five AI-OS Organization Hacks (Nate Herk) — Practitioner rules for organizing the file tree and wikis an agent reads from, framed around the failure they prevent: bad organization makes the agent hallucinate, and the bad context propagates into the skills and automations it builds. The five: routing rules matter, folder shape does not; have the AI audit itself on a cadence; promote repeated just-in-time fetches into scheduled automations; segment knowledge into separate wikis once a node starts growing; and backtrack immediately when retrieval misbehaves. Four of the five map onto machinery this vault already runs — the one that does not is using repetition as the trigger to automate.

Cross-listed: skill-graph plugin packagings

These live under claude-ai/ because they’re Claude Code plugins, but they implement the Karpathy pattern as deliverable software:

• Ars Contexta — Skill Graphs and Conversationally Derived Vault Architecture — Heinrich’s agenticnotetaking/arscontexta (MIT, ~3,300 stars, v0.8.0). 6-phase conversational setup derives a complete self/notes/ops vault + 16 commands (the 6 Rs) + 4 hooks + 249 research claims. Most plugin-packaged version of the pattern. The karpathy-vault is structurally already a skill graph at 270+ articles; arscontexta enumerates three patterns this vault should adopt (promote bin/ scripts to commands, fresh-context-per-phase in /compile, hooks-enforced quality boundary).