Field guide / primary domain
25
sources in this field
Updated August 24, 2026
Current thesis
The shortest path to orientation.
This synthesis records claims and practices from the cited sources; reported outcomes and product capabilities have not been independently verified.
Obsidian remains the dominant IDE for LLM-maintained knowledge bases, with the architecture anchored by Web Clipper ingestion into raw/, agent-compiled markdown with backlinks, and Marp for slides. The community standard is now "vault as foundation, Claude Code as engine"—plain-text markdown that agents read and maintain while humans rarely edit. Vaults split into raw/ (unmodified sources) and wiki/ (agent-maintained knowledge). Retrieval treats the vault like a codebase: an 18-line root index avoids embeddings entirely. Karpathy's key motivation was moving from Notion to Obsidian specifically to use Claude Code and link code context with writing, letting AI act as advanced in-document search across thousands of words instead of requiring elaborate tagging systems.
Vault wiring lives in ~15 lines of global AGENTS.md; outside the vault, agents connect via Local REST API. Two MCP servers bridge agent and vault. Pre-commit hooks and weekly agent passes maintain structural integrity. Tolaria (10,000-note proof, 100K+ LOC, 85% coverage) layers Git and MCP atop plain markdown; ByteRover unifies fragmented notes into relevance-scored indexes. HTML artifacts serve as interactive layers above markdown. Excalidraw (110K stars, end-to-end encrypted) is the de-facto diagram primitive. Google's Open Knowledge Format may displace Obsidian as storage layer while workflow patterns persist.
Evidence board
Claims worth carrying forward01Ian Vanagas draws a hard line between 'writing with AI' (using it for research, questions, and feedback while writing all prose himself) and 'using AI to write' (having it generate final text)—arguing the latter leaves 'skeletons of slop' even after heavy editing.
02He built a Claude 'researcher skill' to find real, quotable, sourced examples because unconstrained prompts caused hallucinated plausible-sounding examples; explicitly demanding sources and quotes keeps the model honest.
03His research tool stack for sourcing high-quality (non-SEO-gamed) examples: Exa (agent-oriented search), Hacker News (target-audience signal), local PostHog repos/RFCs, PostHog Slack, and Semble (a link-network discovery tool)—mirroring the sources he'd use manually.
04Revealed preference: he never reads AI-generated summaries because compression loses the interesting/unique ideas he's actually looking for; he gets more from a couple of good quotes or skimming the source himself.
05AI is a poor editor for tightening prose: it reflects back whatever framing you give it (e.g., it recommended shortening an already-short intro on nearly every review) and struggles specifically with cutting/rewriting tighter, since it's better at adding than removing.
06He moved from Notion to Obsidian specifically to use Claude Code and link code context with writing, letting AI act as 'advanced in-document search' across thousands of words of notes/drafts instead of requiring an elaborate tagging/backlinking system.
07Central unresolved question (from his linked prior post): despite LLMs making developers ~55% faster and enabling viral 'built in 3 hours' coding projects, no equivalent surge of viral 'banger' blog posts has appeared—possibly because writing lacks code's reusable structure ('writing is sand, code is Lego') and writers haven't yet found productized AI workflows the way developers have.
08Case study cited as a research example: PostHog's Wizard agent cost $6.67/run, with a trivial 'conclude' step eating $1.47 due to ~140K tokens of carried context; splitting into fresh query() calls cut input tokens 89% but raised total cost, because Anthropic cache writes cost 12x more than cache reads—so naive context-clearing can backfire economically.
Adjacent fields
Latest evidence
Ian Vanagas
Ian Vanagas distinguishes 'writing with AI' (using AI for research, gap-finding, and fact-checking while keeping prose fully human) from 'using AI to write' (having it draft prose). He details a research skill stack for sourcing real examples, explains why AI is weak at summarization and editing/conciseness, and links this to a broader thesis on why AI hasn't produced 10x more 'banger' blog posts the way it has for code.
Sebastian Kehle
Sebastian Kehle details a ~900-file Karpathy-style Obsidian vault in a private git repo, split into an unmodified raw/ source layer and an agent-owned wiki/ knowledge layer, retrieved via folder indexes instead of embeddings, with automated maintenance hooks keeping it from rotting.
CyrilXBT
A detailed 6-step system for building an Obsidian vault that auto-captures content via Readwise, Airr, Whisper, and a Telegram bot, routes everything through N8N pipelines, and uses Claude to generate daily briefings and weekly syntheses. The core thesis: knowledge systems fail because they optimize for input, not output—feedback loops (daily briefs, weekly synthesis) are what turn a dead archive into a thinking partner.
GREG ISENBERG
Google's open-source Design.md format captures typography, colors, and spacing in a single markdown file that agents reference to produce consistent outputs. Combined with reusable skill files (landing page, mobile, pitch deck), it creates a design system any AI agent can apply uniformly across all surfaces—replacing the common pattern of polishing one screen while everything else looks generic.
Charly Wargnier
Tolaria is a free, open-source (AGPL-3.0) desktop app for Mac and Linux that implements Karpathy's LLM wiki concept — a shared knowledge environment for humans and AI agents. It uses plain markdown, Git-backed vaults, and includes a built-in MCP server for Claude Code integration. Built with Tauri, React, and Rust; 100K+ lines, 85% test coverage, 9.9/10 code health.
Dan McAteer
'Understand Anything' is an open-source plugin for Claude Code, Codex, and OpenCode that analyzes a directory and builds an interactive knowledge base explaining structure rather than showing it. Notably, it generalizes beyond code — running /understand on an 888-episode Obsidian podcast vault (144K lines of markdown) produced a searchable knowledge graph.
Marie Haynes
Google introduced OKF v0.1, an open specification formalizing the 'LLM-wiki' pattern into a portable, interoperable format. A bundle is simply a directory of markdown files with YAML frontmatter — no SDK, no schema registry, no proprietary tooling required. The single required field is `type`; everything else is optional. OKF targets the fragmented organizational context problem: schemas, runbooks, metrics, and join paths scattered across incompatible systems. Reference implementations include a BigQuery enrichment agent and a static HTML visualizer.
Moysei
A step-by-step guide to wiring Claude Code into an Obsidian vault as a self-maintaining knowledge engine, based on Andrej Karpathy's LLM Wiki pattern. The architecture separates immutable sources (raw/) from Claude-compiled linked pages (wiki/), uses MCP + Local REST API for vault access, and schedules daily maintenance at 7am. Key philosophy: start with 10 sources, delegate upkeep to the model, keep human judgment over curation.
Reflect
> Today we’re releasing Reflect Open, a complete rewrite of Reflect for the age of AI. > > It’s local-first and open source, and every note is stored as a Markdown file on your computer. https://t.co/MyFoAsaQaz
Marie Haynes
> This is really big news. Google introduced the Open Knowledge Format (OKF) - a standardized way to store information in a directory of markdown files. Makes it really easy to make a digital brain that agents can use. > > These files can serve as a living wiki. You can give agents the ability to q