Andrej Karpathy tweeted a productivity idea at the beginning of April that has since produced a wave of GitHub repos, YouTube tutorials, and Substack write-ups: use an LLM to build and maintain a personal Markdown wiki from your own source documents. Within hours, the workflow had a name, a following, and a template. Four months later, at least one prominent adopter, Platformer's Casey Newton, says it is the single most useful change to his workflow this year.
Karpathy, who previously coined the term vibe coding, described a simple loop: drop source documents into a local folder, and let an LLM extract, organize, and update a wiki of Markdown files as new material arrives. The system is opinionated about local storage, plain text, and human-readable structure, which makes it durable across model changes and independent of any single vendor's product roadmap.
“Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest.”— Andrej Karpathy, AI researcher
Newton's build sits on top of Claude Code from Anthropic, with Claude Fable 5 generating the initial prompt. The files live in Obsidian, terminal work happens in Ghostty, and queries against the wiki run through an Obsidian plugin called Claudian. The initial seed was six years of Platformer archives going back to 2020, from which Claude extracted the people, companies, and concepts Newton has covered and wrote them up as interlinked Markdown pages.
Key facts
- 01Andrej Karpathy's April tweet on using LLMs to build personal Markdown wikis spawned GitHub repos, YouTube tutorials, and Substack guides within hours.
- 02Platformer's Casey Newton built a wiki of more than 1,440 pages seeded with six years of his own newsletter archive going back to 2020.
- 03The wiki's Meta page alone runs to more than 12,000 words with more than 1,300 internal links to other pages.
- 04The setup uses Claude Fable 5 to generate the initial prompt, Obsidian for storage, Ghostty as terminal, and a plugin named Claudian for queries.
- 05The system replaces a manual 'blips' workflow in Capacities that Newton had used since the summer of 2024 to track long-running stories.
The scale of the resulting knowledge base is the part worth pausing on. The single wiki page for Meta runs to more than 12,000 words and contains more than 1,300 internal links. The full system now spans more than 1,440 pages covering content moderation, child safety, and AI progress — most of what has interested one working journalist across 16 years on the beat.
Each morning, Newton clips a selection of stories into Obsidian, and a local script decides where each one belongs in the existing wiki. New concept pages get generated automatically as they appear in the news — recent examples include tokenmaxxing, youth social media bans, and AI copyright. A home page updates each day with the latest headlines. The workflow replaces a manual system of tracking pages called blips that Newton had maintained in Capacities since the summer of 2024, which collapsed under its own weight as the number of active threads grew.
“Like a vintage sports car, it requires a fair degree of maintenance, and there are almost certainly easier ways to create a personal knowledge base.”— Casey Newton, Platformer founder
The practical payoff shows up in fast-moving stories. Newton points to the OpenAI–Hugging Face agentic breach, which developed in fragments over several weeks, as a case where the wiki paid for itself. Before each podcast or article, he could pull the page for a refresher, then open the original sources to check anything he was about to say out loud. The wiki's timeline structure — every claim links back to a dated source — is what makes it usable for reporting rather than just personal recall.
The setup is not turnkey. Newton's own description compares maintaining it to owning a vintage sports car, and he notes that easier off-the-shelf tools exist. Notion shipped an agent late last year that searches across saved links, and it works, but the search sits on a different surface from the database, and the agent often fails to cite sources without additional prompting. The friction was enough that Newton abandoned the habit. The LLM wiki works because it stays inside one tool — Obsidian — and returns cited answers by default.
The rest of Newton's stack sits alongside the wiki without competing with it. Raycast, now available on Windows as well as Mac, handles launcher duties and runs quick AI searches against GPT-5.5 Instant for the quality-to-speed ratio. Capacities holds the daily journal and tagged Techmeme links. Recall's Chrome extension summarizes YouTube videos, saving hours on podcasts he otherwise would have listened to at 2x. Each of these has survived multiple renewal cycles, which is Newton's real test for productivity software: does it come with you to the next machine.
The genre this fits into is worth naming. Personal knowledge management has been a niche software category for a decade, and the tools have mostly failed to become habits for anyone outside the enthusiast base. What the LLM wiki changes is the maintenance cost. The user still curates inputs, but the system does the writing, linking, and reorganization on its own — which is exactly the kind of tedious, low-stakes, high-context work that current models handle well.
There are limits. Every claim needs verification against the original source, because a wiki page assembled by an LLM will occasionally invent connections that don't exist. The setup also assumes the user is comfortable running scripts, editing prompts, and troubleshooting when the pipeline breaks — which is a smaller audience than the number of people who wish they had a research assistant. And the entire architecture depends on continued access to models with strong long-context extraction, which for now means paying Anthropic or a competitor per token indefinitely.
What Karpathy identified, and what Newton's build demonstrates, is that the useful unit of AI-assisted knowledge work may not be the chatbot query. It's the durable, local, plain-text artifact that a model helps you maintain over years. That framing has implications for every productivity vendor currently trying to bolt an assistant onto a proprietary database. If the winning pattern is Markdown files on a laptop, queried by whichever model is best this quarter, the moat belongs to the format, not the app. Notion, Capacities, and the rest have a decision to make about whether to embrace that or fight it.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
