Make Your IG Saves Useful

Grab it from obsidian.md, open it, hit Create new vault, name it, and pick a spot on disk.

Under the hood, a vault is just a folder of markdown files. That's the whole storage layer. No database, no lock-in, and your files stay yours.

Open a terminal inside your vault, launch claude, and hand it this:

Set up my second brain in this folder:

- raw/ is my inbox. I'll dump unprocessed material in here: articles, links, transcripts, notes. No sorting, no tags. I drop it and forget it.
- wiki/ is yours. You'll write and maintain everything inside it. Create wiki/_master-index.md with the heading "Knowledge Base" and the line "Topics appear here as they're created."
- output/ is where generated reports and answers land.

Create just the folders and that one file. Nothing else yet.

Resist the urge to add structure on day one. If you want PARA folders or fancy templates later, layer them in when the vault outgrows what you have. Fewer rules, faster start.

One file does almost all the work. It holds the rules, the workflows, the house style, the folder layout, and how the brain should evolve over time. Claude reads it at the start of every session. Get this right and the rest runs itself.

It's also the honest caveat of this whole system: the agent is only as good as these instructions. Vague rules make vague notes. When something it writes bugs you, fix the rule here, not just the note.

# CLAUDE.md - my second brain

You are the librarian of this vault. Read this file at the start of every session.

## How this vault works

- raw/ is the inbox. I drop unprocessed material there: articles, transcripts, links, voice notes. I never sort it.
- wiki/ is yours. You write and maintain everything in it. I read it, but I rarely edit it.
- wiki/_master-index.md is the front door: every topic, one line each. Keep it current.
- Every topic gets its own folder (like wiki/ai-tools/) with an _index.md listing its articles.
- output/ is where reports and query answers land.
- Cross-link aggressively with [[wiki links]].

## On "compile"

1. Read each file in raw/.
2. Pick the right topic, or create a new one.
3. Write a concise wiki article: key takeaways first, then supporting detail, then links.
4. Update the topic _index.md and the master index.
5. If one raw file spans several topics, split it and cross-link.
6. Move processed files to raw/done/. Never delete anything without asking.

## House style

- Bullets over paragraphs. Short and concrete.
- Every article ends with a ## Key Takeaways section.
- Filenames: lowercase-with-hyphens.md
- Keep my voice when cleaning up my own notes.

## On questions

Route through the indexes: _master-index.md, then the topic _index.md, then the 1-3 most relevant articles. Synthesize from those. Don't scan the whole vault.

## On "audit"

Walk the wiki. Report inconsistencies, broken links, gaps, and topics that are referenced but have no article of their own. Report only, don't edit.

## How this file evolves

When I correct you twice about the same thing, propose a new rule for this file.

The last rule is the sleeper. When you correct the agent twice about the same thing, it proposes a new rule for its own instructions. The system tunes itself to you over time.

The system only works if saving something takes one click. Install the Obsidian Web Clipper extension in your browser, then in its settings: point Vault at yours, set Note location to raw, and leave the default template alone.

Now any article is one click into your inbox as markdown. Video links, PDFs, voice memo transcripts: same folder, just dump them. On your phone, the Obsidian mobile app plus the share sheet does the same job.

Optional but worth it: the Local Images Plus community plugin pulls images from clipped pages into your vault, so nothing rots when a source page disappears.

Clip 3 to 5 articles on a single theme first, so the first run has something to connect. Then tell Claude:

Compile everything in raw/. For each file: read it, pick or create a topic in wiki/, write a tight article with key takeaways, add [[wiki links]] to related ideas, then update the topic index and the master index. Cross-link aggressively. When you're done, move the processed files to raw/done/ and give me a one-line summary per article.

Refresh Obsidian. The agent went into the inbox, extracted what mattered, and turned your messy dump into structured, connected notes. Topic folders, indexes, articles linked into a graph. Open the graph view if you want the dopamine hit.

Check the first few compiles before you trust it. Read what it wrote, and when the tone or the filing is off, add a rule to CLAUDE.md. Ten minutes of tuning early saves you a messy vault later.

Three flavors, in rising order of magic. Direct lookup: "what were the main findings from that article on X?" Cross-reference: "pulling from the whole wiki, how do A and B relate?" And synthesis, the one worth copying:

Based on everything in the wiki, what are the main approaches to [topic] and how do they trade off against each other? Save your answer as a new wiki article, link it to the sources you pulled from, and update the indexes.

Synthesis is the loop that compounds. The answer gets filed back into the wiki, so the next question has more to work with. The system gets denser every time you use it.

And notice what's missing: no vector database, no embeddings, no API bill for search. The indexes do the routing. Claude reads the master index, then the topic index, then the 1 to 3 relevant articles. Three or four targeted reads per question, and it scales to thousands of files.

This is the part most people miss. You don't maintain the system when you remember. You maintain it on a schedule, so it stops depending on your discipline.

Wire these up however you like: calendar reminders, cron, or scheduled Claude Code runs. The point is that you're no longer the one deciding when to maintain it.