🎯 The Problem: Second Brains Are Too Complicated 0:00
The observation that drives the whole video: half of the AI second-brain setups don't actually work, and the other half are way too complicated for 99% of people. It's easy to install someone else's system and "not have a clue about what's going on." Chris Ashby (15 years designing apps and advising startups) argues you don't need to start with Obsidian, with its interlinked knowledge graph that looks impressive but is a maintenance burden — you can start with a simple folder structure and markdown files that an AI agent reads and curates.
🔮 Karpathy's Idea: The Self-Documenting Knowledge Base 3:34
SimpleBrain is "built on top of Andrej Karpathy's idea for a self-improving, self-documenting knowledge base" — the viral post that got 21M+ views. The three pillars, distilled to their simplest form:
- A raw folder where you dump everything readable by an agent — PDFs, markdown, CSVs, JSON, website URLs.
- The AI builds the brain — automatically documenting your raw files (on a schedule or on trigger) instead of you maintaining it yourself.
- Extend with automations — new folders, prompts, and tools layered on top, or use it as a thought partner for projects and business.
📁 The Structure: 5 Folders + 3 Files 6:39
"Five folders and three files. That is literally it."
| Item | Purpose |
|---|---|
raw/ | Dump everything — PDFs, markdown, CSVs, JSON, URLs, documents. |
wiki/ | The knowledge base the agent curates from raw. |
archive/ | Old raw files after they've been curated. |
prompts/ | Reusable prompts for automations (e.g. translate.md). |
projects/ | Individual projects, when used as a business second brain. |
CLAUDE.md / AGENTS.md | Instruction files the agent runs before your prompt — goals, about you, tone of voice (can start blank and let the AI fill them). |
README.md | The loop, in one line: "dump raw stuff into raw, AI turns it into clean notes in wiki." |
The flow is dead simple: AI takes files raw → wiki (builds the brain), and raw → archive (archives processed files). The whole folder is a Git repo, so nothing is ever lost — and you can keep the repo on GitHub so a team shares one source of truth in the cloud.
⚙️ Setup in Claude Co-work 8:49
Setup is just: download the SimpleBrain folders from GitHub, unzip, then in Claude Co-work create a new project → use an existing folder → select the SimpleBrain folder → create. You immediately get the folder structure and context visible in the desktop app. (The video is sponsored by Anthropic, and the host is explicit about it — he uses Claude Co-work daily for exactly this.) The same setup works with almost any agent tool — Codex, Cursor, or any agent that reads a CLAUDE.md/AGENTS.md.
📜 The Rules: CLAUDE.md, AGENTS.md, translate.md 13:40
The instruction files are where the safety lives, and they're designed to evolve with you. The common tasks: translate raw → wiki, project digest (summarize a project into its README), and answer questions (read the wiki and archive to answer about your past thinking). The rules, worth copying verbatim:
- Never delete anything from raw or archive.
- Never overwrite the wiki blindly — always read first, then merge.
- Never modify archive files — that's old, settled material.
- Commit to Git after anything meaningful changes — a save point.
- When uncertain, log it in the entry rather than guessing.
The translate.md prompt is the engine: read everything in raw, update or create matching markdown files in wiki so the wiki reflects raw, then move each processed file to archive. Rules: one topic per file, merge into existing entries, link related entries, never delete, surface to the relevant project's notes if it belongs to one, and note uncertainty instead of guessing.
🎬 The Demo: Raw → Wiki in Minutes 16:26
The demo makes it concrete: he asks Claude Co-work to read his business website and capture everything about it into a markdown file in raw. Then a single instruction — "run the translate prompt in the prompts folder" — and the agent reads the translate.md prompt, the README, and the raw file together, and produces a whole wiki: separate markdown files for the brand, the community arm, and the individual product offerings, each linking to the others. "In just a few minutes we're getting set up with a complete knowledge base of our business, just from raw information dumped in the raw folder." The processed raw file lands in archive, ready for the next drop.
⏰ The Daily Automation 20:50
You don't want to run the translate prompt by hand forever. In Claude Co-work you set up a scheduled task: "daily translate," working in the SimpleBrain project, running the translate prompt every day at 9 a.m. The model choice is deliberate — Sonnet 4.6 rather than a flagship, because translating raw notes into wiki entries doesn't need peak intelligence and the cheaper model "helps cut down costs." That single automation is the moment the second brain stops being a thing you maintain and starts growing on its own.
🧩 Extensions: Voice, YouTube, Weekly Synthesis 23:08
From there, the automations stack:
- Voice memo → transcript. Record voice notes on your phone while walking the dog; a watched folder transcribes them with Whisper into raw, and the daily translate sends them to the wiki.
- YouTube / podcast transcripts. Check a channel daily or weekly, pull new videos, transcribe them with yt-dlp, and feed the transcripts through the same raw → wiki pipeline.
- Weekly synthesis. A prompt that reads the last week's wiki changes and summarizes the themes, decisions, and open questions — a standing digest of your own evolving knowledge.
You can run one second brain per business, or one per project, and the simple starting point is what makes them "way less difficult to maintain over time" than the complex systems everyone else demos.
💡 Key Takeaways
- You don't need Obsidian. A markdown editor and a knowledge graph are the end state, not the starting point.
- Five folders, three files. raw → wiki → archive, plus prompts and projects — that's the whole system.
- Let the AI build the brain. You dump raw files; the agent documents them into a curated wiki on a schedule.
- Encode safety in the instruction files. Never delete, merge before overwriting, never touch archive, commit to Git, log uncertainty instead of guessing.
- One prompt is the engine. translate.md — read raw, update wiki, archive the processed file — one topic per file, linked entries.
- Automate it daily on a cheap model. A scheduled task on Sonnet 4.6 does the translation every morning at a fraction of the cost.
- It's a Git repo. Nothing is ever lost, and GitHub turns it into a team-shared source of truth.
- Extend, don't rebuild. Voice memos, YouTube transcripts, and weekly synthesis all layer onto the same raw → wiki loop.
🔗 Resources & Links
- 🐙 github.com/BuildGreatProducts/SimpleBrain — the folder/file setup to duplicate
Source video: youtube.com/watch?v=h0HYpONXgjk (sponsored by Anthropic, stated in the video). Built on Andrej Karpathy's self-documenting knowledge-base concept.