The Team Brain · Module 2: The knowledge
Structure for a machine
AI tools find answers better in documents that are structured for retrieval. This does not mean anything technical. It means: one topic per document, a clear title that matches how people ask, headings that are questions, short paragraphs, the answer near the top, and a date on everything.
Long documents covering many topics should be split. A forty-page handbook becomes twenty two-page documents, each titled by the question it answers. A table of contents document, listing every document and what it covers, helps too, both for the machine and for humans.
Then the folder. One place, with a clear name, that the assistant reads from and nothing else. If you use SharePoint or Google Drive, that is the folder you connect. If you use a tool that takes uploads, that is the set you upload. Keeping it to one place is what makes the next module's maintenance routine possible.
| Do | Avoid |
|---|---|
| One topic per document, titled as the question | Long handbooks covering everything |
| Answer in the first paragraph | Answer buried after background |
| Headings phrased as questions people ask | Headings like "Section 4.2" |
| A date on every document | Undated documents nobody trusts |
| One folder the assistant reads from | Documents scattered across drives |
In practice. A team split its 60-page operations manual into 34 short documents, each titled with the question it answered, using AI to do the splitting and a human to check the titles. Search improved for people before any assistant existed. The assistant, when built, cited the right document first time in almost every test.
At your desk. Take your current folder. Split anything covering more than one topic. Retitle every document as the question it answers. Date everything. Write the table of contents document.
Write it down.
- Documents in the final set:
- The document I split into most pieces:
- The folder the assistant will read: