Connecting the brain gives your AI the tools. It does not tell the AI when to use them. A user can finish sign-in, see the connector listed, see every tool enabled, and still get answers from the model's own memory while a fully connected brain sits untouched. What fixes this is one short block of text in the client's custom instructions. This page explains what the block does, why it is short, and where it goes in each client.
The block
The dashboard generates it for your account under Your brain, then Connect AI client, already listing every brain you can reach. For one brain it reads:
My LAN Brain connector (Acme) is the source of truth about me and my work.
Before answering anything that depends on my context (my projects, my company, my decisions, the people I work with), call search_brain first. If the brain has nothing, say so instead of guessing. When the brain conflicts with what you think you know, the brain wins; cite the page path.
Do NOT use the connector for general knowledge: definitions, public facts, coding help, anything you could answer without knowing me.
Answer bottom-line-up-front: the answer first, then only the context that changes my next step.
When I correct you on something durable, fold it into the brain (apply_learning) and tell me in one line what you updated. A single ambiguous comment is not a durable correction.
Each line earns its place:
| Line | What it prevents |
|---|---|
| Source of truth, search first | Answers from memory about your company, your clients, your decisions |
| Say so if the brain has nothing | Confident guesses dressed up as facts |
| Brain wins, cite the path | Stale model knowledge overriding a decision you wrote down last week |
| Do NOT use it for general knowledge | The AI reaching for the brain on every question, which is slow, wastes tokens, and is the fastest way to get the connector switched off |
| Bottom line up front | Long answers that bury the point |
| Fold corrections into the brain | Telling the AI the same thing every week |
Why it is short
Your AI reads custom instructions on every single message, including chats that never touch the brain. A long block is a tax you pay every time you type. So the block carries only what must exist before the AI ever talks to a brain: the triggers and the routing. Everything else (the list of tools, how skills work, how to file a page, how to route to your CRM or project tracker) is delivered by the brain itself at the start of each session, and by each tool's own description at the moment it is called. That is why the block stays under about 150 words and why you should not add to it.
More than one brain
If you have a personal brain and one or more company brains, the dashboard generates a single block with one routing line per brain:
My LAN Brain connectors are the source of truth about me and my work. Routing:
- Isaac (personal): me, my life, my own projects.
- Acme: anything about Acme's business, people, and decisions.
Pick the brain that owns the subject; a company's own facts beat my personal notes about it. ...
One block covers all of them. It replaces any earlier LAN Brain block you pasted, so paste once and delete the old one.
Where to paste it
| Client | Where |
|---|---|
| Claude (web, desktop, mobile) | Your initials, then Settings, then Profile, the Instructions for Claude field (older builds: personal preferences). Account-wide |
| Claude Code | ~/.claude/CLAUDE.md for every project, or ./CLAUDE.md in the repo to share with the team |
| ChatGPT | Settings, then Personalization, then Custom instructions (enable customization) |
| Gemini app | Saved info at gemini.google.com/saved-info |
| Gemini CLI | GEMINI.md in the project or in ~/.gemini/ |
| Cursor | Settings, then Rules, then User Rules; or .cursor/rules/company-brain.mdc with alwaysApply: true |
| VS Code and Copilot | .github/copilot-instructions.md in the repo, or a user instructions file |
| Windsurf | Global rules under Cascade settings, or .windsurfrules |
| Zed | .rules at the repo root, or the agent profile instructions |
What not to add
- Do not paste tool lists or protocols. The brain sends them itself; duplicating them makes the block longer and can go stale.
- Do not add "always use the brain". The negative clause is there on purpose. An AI that searches the brain to define a word gets turned off within a week.
- Do not keep two versions. If you had an older LAN Brain block, delete it when you paste the new one.
- Do not translate it. The block is an instruction for an AI, not prose for you; it stays in English in every client.
How to know it is working
Ask something only your company knows and look for a page path in the answer, such as wiki/decisions/d-012-enterprise-pricing.md. A path means the AI searched. No path means it did not; say "check the brain first" once and confirm the block is saved in the right place. If it still does not search, see Troubleshooting.