The idea in one sentence

Imagine a ChatGPT plugged into your own documents, living in a folder on your computer.

You drop in whatever you want: meeting notes, operating procedures, quotes, customer complaints, exported e-mails, memos. The assistant reads them, summarizes them, extracts what matters, creates reference sheets, spots contradictions, and keeps everything organized. When you ask a question, it doesn't start over: it consults the base it has already built.

Honest note: your documents stay on your disk, but when you ask a question, their content travels encrypted to Anthropic's servers to generate the answer. This is not a 100% local system. It is, however, a system where you keep control of the perimeter: nothing leaves the folder without your action.

Why it beats a classic upload

With ChatGPT, NotebookLM or "Projects", the AI rediscovers your documents in every conversation. Nothing accumulates. The system we install here does the opposite: with each new source, the AI updates a persistent knowledge base. It enriches existing sheets, flags what's new, cross-references. The system gets more useful as it grows, like a library someone tidies for you.

This pattern is called LLM Wiki, formalized by Andrej Karpathy in 2024.

The three zones of the system

That's all. No database, no server: a folder of text files.

What you will install

  1. A working folder on your disk · somewhere stable, ideally synced to your cloud (automatic backup). For example Documents/projects/company-memory, no spaces, lowercase.
  2. Obsidian · the free editor that gives you a comfortable view of the folder: outline, clickable links between sheets, graph view. On first launch, choose "Open folder as vault" and point it at your folder.
  3. Claude Code Desktop · the Anthropic app that can read and write in a local folder. It plays the librarian. Sign in to your account, then select the same folder.
  4. Obsidian Web Clipper · the browser extension that turns any web page into a file in sources/. In its settings, change the destination folder from "Clippings" to "sources".

Prerequisites: an Anthropic account with a Pro plan (€20/month) or higher, and install rights on your computer.

The first conversation

Your folder is empty; ask Claude to prepare it. Paste this into Claude Code Desktop:

Ok, I would like to start by installing llm-wiki · https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f · in this folder. The goal is to drop documents in (sources/) and let you maintain synthetic reference sheets (wiki/) with clear instructions (CLAUDE.md). Create the folder structure and a CLAUDE.md suited to an SME [describe your business]. We will then start adding sources.

Claude reads the reference, proposes a structure, asks a few questions, creates the folders and steering files. For each file it creates, it asks permission: that's intentional and healthy · nothing happens without your validation.

Then open CLAUDE.md in Obsidian and adapt it to your trade: the sheet categories of a machine shop are not those of a consulting firm. The AI re-reads this file at every session: your changes apply immediately.

First cycle: one source, one question

  1. Capture a source · a trade article, a supplier announcement, a technical doc page · in one click with the Web Clipper. The file lands in sources/.
  2. Have it ingested:

Ingest the file sources/[file-name].md following the workflow defined in CLAUDE.md.

Claude reads, discusses the key points with you, proposes the sheets to create, updates the catalog and the journal. One to two minutes per source · which you can watch appear live in Obsidian.

  1. Ask a question:

What are the key points you retain from what you just ingested? Cite your sheets.

You just completed a full cycle: source → ingestion → sheet → question → synthesis. That's the gesture you will repeat hundreds of times, and with each repetition the base gains value.

What now? Four tracks for an SME

The owner first. Start with your own knowledge: meeting notes, decisions, contacts, deal history. Within weeks you have a queryable company memory.

The shop floor. Operating procedures, recurring breakdowns and their fixes, machine settings: the knowledge that retires with the veterans goes into the wiki.

Quality. Every non-conformity dropped into sources/ enriches the sheets: recurring causes, suppliers involved, actions that worked.

Marketing and customer service. The base knows your products, your customer cases, your vocabulary: it becomes the raw material for accurate content, and fast answers.

In every case, the rule is the same: be the curator, let the AI be the librarian. You decide what goes in; it reads, summarizes, cross-references, files and updates.

To go even further, discover @Optimike's method: a complete Obsidian course, in French.