It is a set of markdown pages that an AI model writes and maintains from your source material. Instead of searching raw documents for every question, the model reads each new source once, merges what it learned into concept and entity pages, and keeps links and summaries up to date.

How do I set up Claude Code with Obsidian for notes?

Make an Obsidian vault with a read-only folder for raw sources and a folder the AI owns for wiki pages. Add an index file listing every page, a log file, and a CLAUDE.md at the root describing the rules. Then ask Claude Code to ingest new sources and update the index.

Is an LLM wiki better than NotebookLM?

They solve different problems. NotebookLM answers questions over the files you load into a notebook. An LLM wiki builds lasting pages you can browse and edit, which suits long-running topics, but it needs upkeep and review because the model writes the pages.

All video notes

Video notes · Notes and PKM

Complete Guide to building an AI Second Brain with Claude

Video by Ray Fu on YouTube · 12 min · uploaded October 2026

Notes and analysis by BrainTube, written from the video. Not affiliated with Ray Fu. Watch on YouTube

In short

Ray Fu walks through an LLM wiki in the style Andrej Karpathy described: Claude Code reads a read-only folder of raw sources and maintains a linked markdown wiki in Obsidian, steered by a CLAUDE.md schema and three identity files. The folder rules and the index-first lookup are sound. But the headline accuracy claim has no method behind it, the key prompts sit behind a link, and the habit of saving AI answers back into the wiki can compound errors unless sources stay the ground truth.

The argument

The video argues that tools like NotebookLM, file uploads and plain vector search rebuild an answer from fragments every time, so nothing accumulates. An LLM wiki fixes that by compiling knowledge once: each new source is read, its entities extracted and merged into persistent pages, and the model keeps cross-references and summaries current. He frames it like software, with raw notes as inputs, Claude Code as the programmer, the wiki as the codebase and Obsidian as the IDE. Strict folder boundaries, a CLAUDE.md schema and identity files make the output accurate and personal, and daily automations keep it growing with little manual effort.

Key ideas, with timestamps

2:10
Why plain retrieval does not accumulate
In NotebookLM, ChatGPT file uploads and standard vector databases, the model retrieves chunks and synthesises an answer per question, then the synthesis is gone. A question that needs five documents from the last three months gets pieced together from scratch every run.
2:43
Compile once, keep updating
Instead of searching raw files at query time, the AI maintains a persistent wiki. Adding a source means reading it, pulling out the core entities and merging them into existing pages. The tedious upkeep, such as cross-references and summaries, becomes the model's job.
4:22
Raw sources are read-only
The vault has a raw sources folder for articles, PDFs, transcripts, screenshots and papers, with a rule that the AI may read but never edit or delete them. He claims accuracy went from 60% to 95% once he added structural boundaries like this, without saying how it was measured.
4:55
The wiki, index.md and log.md
The wiki folder is entirely AI-owned: concept pages, entity pages and comparisons. index.md lists every page with a link and a one-line summary by category, and Claude reads it first on every query. log.md is a timestamped change log of ingests, queries and health checks.
6:02
Three identity files
user.md says who you are and how you like to work. soul.md defines the assistant's role; his asks for a direct, critical chief of staff that challenges assumptions. anti-AI-style.md is a ban list of phrases he dislikes. At 7:09 he suggests having Claude interview you by voice for 15 minutes to draft them.
7:41
Feeding it with the Web Clipper
The Obsidian Web Clipper extension, pointed at the raw sources folder, saves articles in one click. A single ingest command then has Claude Code create wiki pages, update the index, write the log and link concepts with Obsidian wiki links.
8:46
Index first, then save answers back
Queries start at index.md, move to the relevant cluster and follow wiki links, with citations in the answer. His rule for compounding value: when Claude produces a good comparison or breakdown, have it saved as a new wiki page and added to the index.
9:50
Three automations and backups
After a community plug, he adds a 7:30 morning briefing on a cron job, meeting transcripts ingested into the wiki, with a note-taker MCP to sync them, and a weekly health check to catch wrong statements. At 10:55 the Obsidian Git plugin commits the vault to a private GitHub repo every 15 minutes.

What it gets right

The read-only rule for raw sources is the most important design choice in the video, and he states it clearly. As long as the originals are untouched, any wiki page can be checked or rebuilt. Splitting what the AI may write from what it may only read is good practice for any agent working on your files.

The index-first lookup is also sound for a wiki of modest size. A one-line summary per page lets the model choose what to open without reading everything, and a log of changes gives you a way to see when and why a page shifted.

The weekly health check and git backup are the kind of maintenance most tutorials skip. A knowledge base written by a model needs review, and version history lets you undo a bad ingest.

What it leaves out

The accuracy numbers. The jump from 60% to 95% at 4:22 comes with no test set, question count or scoring method. Treat it as an anecdote. If you build this, write 20 questions with known answers and rerun them after changes.
Saving answers back can compound errors. At 8:46, model-written analysis becomes a wiki page that later answers build on. If one comparison is wrong, it gets cited again. Mark saved answers as derived, link them to the raw sources they used, and let the health check recheck them against those sources.
The schema is gated. The CLAUDE.md prompt that defines how pages are built, which drives most of the result, is only in a guide behind a link. The video shows the folder layout but not the rules that make it work.
Index limits. A single index.md read on every query works for hundreds of pages. At thousands it grows too long to read cheaply, and you need sub-indexes or search.
Not everything should be compiled. Turning every meeting and inbox item into wiki pages adds pages that go stale. Some material is better fetched when needed than merged in.
Privacy. Meeting transcripts and client material are sent to the model provider during ingest. That deserves a decision before you automate it. Roughly two of the twelve minutes are giveaway and community promotion.

How this applies to memory across AI tools

The compile-once idea is the right one: process a source a single time into a form you can query, rather than re-reading raw files for every question. The limit of this build is reach. The wiki lives in an Obsidian folder that Claude Code or Codex can read on one computer; a chat assistant on another device cannot see it. See persistent AI memory across tools for that problem.

BrainTube applies the same compile step to the sources this video clips by hand: it saves YouTube videos with their transcripts, articles, PDFs and podcasts into one compiled knowledge base and exposes it as an MCP server, so Claude, ChatGPT, Cursor and other MCP clients can search it. Whichever route you take, keep the originals as the ground truth. For how this compares with notes apps, see personal knowledge base for AI.

Who should watch it

Watch it if you want a quick picture of an LLM wiki in Obsidian and are willing to write your own CLAUDE.md rules. The useful stretch is 2:10 to 9:18: the case against chunk retrieval, the folder rules at 4:22, the identity files at 6:02 and the save-back rule at 8:46. Skip 0:00 to 1:36, which is mostly introduction and a giveaway, and 9:18 to 9:50, a community pitch. Skip it entirely if you already run a wiki like this or want measured results; the video gives a layout, not tested prompts.

Frequently asked

What is an LLM wiki?: It is a set of markdown pages that an AI model writes and maintains from your source material. Instead of searching raw documents for every question, the model reads each new source once, merges what it learned into concept and entity pages, and keeps links and summaries up to date.
How do I set up Claude Code with Obsidian for notes?: Make an Obsidian vault with a read-only folder for raw sources and a folder the AI owns for wiki pages. Add an index file listing every page, a log file, and a CLAUDE.md at the root describing the rules. Then ask Claude Code to ingest new sources and update the index.
Is an LLM wiki better than NotebookLM?: They solve different problems. NotebookLM answers questions over the files you load into a notebook. An LLM wiki builds lasting pages you can browse and edit, which suits long-running topics, but it needs upkeep and review because the model writes the pages.

Keep videos like this one searchable

BrainTube saves YouTube videos with their transcripts into one compiled knowledge base that Claude, ChatGPT, Cursor and other MCP clients can search.

Start free

Complete Guide to building an AI Second Brain with Claude

In short

The argument

Key ideas, with timestamps

Why plain retrieval does not accumulate

Compile once, keep updating

Raw sources are read-only

The wiki, index.md and log.md

Three identity files

Feeding it with the Web Clipper

Index first, then save answers back

Three automations and backups