Obsidian + Claude second brain that still works at 1,000 notes

Most "Obsidian + Claude second brain" guides stop at "connect Claude to your vault". This one covers what happens after: four folders, one short index page, Claude-written topic pages that cite their sources, and a weekly check so the wrong stuff gets fixed. Gathered from Andrej Karpathy's "LLM wiki" note, Anthropic's research on agent memory and the Claude Code docs, rewritten in plain words.

Start here

How to use this canvas

Top to bottom: why notes get lost, the 4-folder layout, how Claude fills it, then the weekly check.
You can start with 10 old notes. You don't need to move your whole vault.
Works with Claude Code (reads the whole folder) or the Claude app (you upload a few pages). Both are covered.

Why notes get lost

Saving is easy. Finding is hard. After a few hundred notes, search returns dozens of half-matches and you read none of them.
The fix isn't more tags. It's a layer on top: short topic pages that summarise many notes, and one index that lists the topic pages.
Andrej Karpathy calls this an "LLM wiki": you collect sources, the AI writes and maintains the wiki pages, you read them (gist, 2026-04).

What you'll end with

A vault with 4 folders and a rule file Claude reads first.
One index page you can read in 30 seconds.
3 topic pages written from 10 old notes, each with links to its sources.
A 10-minute Friday check that catches wrong or duplicate pages.

The layout

Four folders. That's it.

| Folder | What goes in | Who writes it |
|---|---|---|
| Inbox | Anything new: clips, quick notes | You |
| Sources | Originals you keep: articles, transcripts, PDFs | You. Claude only reads these |
| Topics | One page per topic, summarising the sources | Claude |
| Index | index.md (the short list) and log.md (what changed, when) | Claude |
Obsidian treats a vault as a normal folder of Markdown files on your computer, so Claude can work on it like any folder. Save web articles straight into Sources with the official Obsidian Web Clipper.

Keep the index short on purpose

index.md: one line per topic page, with a link and a one-sentence summary. Nothing else.
Why short: Anthropic's research says the layer read every session has to stay relatively small. Claude Code loads only the first 200 lines (or 25 KB) of its own memory index for the same reason (2026-11).
Detail lives in topic pages. The index only points to them.

The rule file Claude reads first

Put a file called CLAUDE.md in the vault's top folder. Claude Code reads it at the start of every session. Ask Claude to draft it:

Write a CLAUDE.md for my notes folder. Rules: Inbox and Sources are mine, never edit or delete them. You write pages in Topics, one topic per page, and every claim ends with a link to the source note it came from. Keep Index/index.md to one line per topic page. Add one line to Index/log.md each time you change something: date, what, why. Under 40 lines total.

Claude fills it

First run: 10 notes into 3 topic pages

Read Index/index.md, then the 10 newest notes in Inbox. Group them into at most 3 topics. For each topic, write a page in Topics: a 5-line summary at the top, then the key points, each ending with a link to the note it came from. Move the notes from Inbox to Sources. Update the index and the log. Show me the plan before you write anything.

"Show me the plan first" is the important line. Read it, fix it, then say go.

Every topic page says where it came from

Each point ends with a link like [[2026-11-03 pricing call]]. In Obsidian that's a clickable link to the source note.
If a point has no source, it shouldn't be on the page. That's how you stop the wiki from slowly making things up.
Karpathy's version also files good answers back in: when Claude answers a question well, save that answer as a topic page too.

Ask questions the right way

Answer from my notes only. Start from Index/index.md, open the topic pages you need, and quote the source notes. If the notes don't cover it, say so instead of guessing.

Starting from the index keeps Claude from opening hundreds of files every time.

Keep it right

Fridays: check it against the sources

A wrong page gets trusted by every later answer. In one of Anthropic's experiments, Claude checked a note against the original session transcript and caught a wrong one. They also say checking all memory all the time might not be practical, so we check once a week (2026-11). A small manual version:

Check every page in Topics that changed this week against the source notes it links to. List: claims that disagree with their source, claims with no source, pages that cover the same topic, and links that point nowhere. Don't change anything yet.

Read the list, then tell Claude which fixes to make.

One topic, one page, one owner

Two pages about the same thing drift apart. Merge them.
Anthropic saw facts get lost when several writers edited one shared file; one page each fixed it. For you: Claude writes Topics, you write Inbox and Sources, nobody edits the other's.
Fewer, better pages beat many pages. In the same research, a run that wrote about 10 memory files got further than one that wrote 413. (The two runs used different models, so this shows a direction, not a strict comparison.)

Two ways to connect Claude / Honest bit

Claude Code: reads the whole folder

1. Install Claude Code from Anthropic's docs.
2. In a terminal, go to your vault folder and type claude.
3. It reads CLAUDE.md, then works on the files directly.
Best when you have hundreds of notes. Some guides add the Obsidian CLI or an MCP plugin on top; you don't need them to start.

Claude app: upload a few pages

1. Make a Claude project called "Notes".
2. Paste your CLAUDE.md rules into the project instructions.
3. Upload index.md and the topic pages you're working on.
4. Copy Claude's new or fixed pages back into the Topics folder yourself.
Slower, but no terminal. Fine for under about 50 topic pages.

Limits worth knowing

Claude can only organise what you give it. If a note never reaches Sources, the wiki won't know it.
It can still write something wrong with a source link next to it. The Friday check is how you catch that.
Don't paste whole paid articles or books into Sources: save the link and your own summary.
Obsidian is free to download; a commercial license is optional. Claude Code needs a paid Claude plan or API credits. Check current prices before you start.
We organised this from the tutorials and docs in Sources. We didn't test the MCP plugins.

Try it this week / Sources

Do the 10-note version today

Make the 4 folders, add CLAUDE.md, put 10 old notes in Inbox and run the first-run prompt. On Friday, run the check once.

Pair it with "Make Claude remember you"

Your notes wiki is what you know. Claude's memory is who you are and how you work. The "Make Claude remember you" guide covers that half: https://puthere.space/collections/make-claude-remember-you-en?ref=guide-obsidian-second-brain

Plan the vault on one canvas

Sketch your 4 folders, your topic list and your Friday checklist as cards on one put·here canvas before you move any files. Download put·here for Mac: https://puthere.space/?ref=guide-obsidian-second-brain · iPhone: https://apps.apple.com/app/id6823383945?ct=guide-obsidian-second-brain

Sources

Andrej Karpathy, "llm-wiki" (third party; GitHub gist, 2026-04). Raw sources, an AI-maintained wiki, a schema file such as CLAUDE.md, index.md and log.md, ingest, query and lint. https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
Anthropic, "The components of agent memory" (official research; 2026-11). Small top layer, checking notes against transcripts, one page per writer, lean memory. https://www.anthropic.com/research/components-of-agent-memory
Claude Code docs, memory (official; checked 2026-11). CLAUDE.md read each session; first 200 lines of the memory index. https://code.claude.com/docs/en/memory
Obsidian Help (official; checked 2026-11). Vaults, Web Clipper, commercial license. https://obsidian.md/help
NxCode, "Obsidian AI Second Brain: Complete Setup Guide" (third party; updated 2026-11). Ways to connect Claude: Claude Code in the vault, Obsidian CLI, MCP. https://www.nxcode.io/resources/news/obsidian-ai-second-brain-complete-guide-2026