Build a living knowledge base

Step 1 — Pick the domain you keep re-explaining

The tell is repetition. Whatever you type into every new chat — “context: I run a…” — that’s your domain. One domain only; small and real beats big and abstract. The goal of this lab is that you never type that preamble again.

Step 2 — Let the AI interview you

Nobody can write their own documentation from a blank page. Everybody can answer questions. So don’t write — be interviewed:

I’m building a knowledge base about [domain] that future AI conversations — and colleagues — will work from. Interview me: one question at a time, starting with what a total stranger would need first, then the details that actually matter day to day. When you have enough for a first version, say so.

Answer in fragments; ramble; use voice input if you have it. The interview is extraction — it’s how the knowledge gets out of your head and into a window for the first time.

Step 3 — Structure it hot and cold

Organize everything I told you into a knowledge base with two tiers. THE CORE — one page, maximum: what this is, who’s involved, the current goal, the current state, and where everything else lives. THE SHELF — separate pages by topic, each self-contained, each opening with a one-line summary. Write every page for a reader with zero context. Flag anything I said that’s missing a piece or contradicts itself. Put “Current as of [today’s date]” at the top of every page.

This is the hot/cold pattern from Source of truth — play the hot–cold toy there if you haven’t. The core is what every future window carries; the shelf is what it reaches for. A knowledge base that’s all one giant page is a rotting window waiting to happen.

Step 4 — The acceptance test: a stranger window

Open a fresh chat — no history, no memory of the interview. Paste the core (and one relevant shelf page), then ask a real question you’d normally need ten minutes of preamble for:

Using only the knowledge base below, [do the real thing — draft the reply to this client, plan next week, decide between these two options]. Then tell me what the base didn’t contain that you found yourself needing.

If the stranger window acts correctly, the base works — that’s the whole test. Not “is it complete,” but can someone with zero context act from it. Whatever it flagged as missing goes in now, while you remember the answer.

Step 5 — Make it live where the work happens

Move it into its home — Notion, Docs, the folder — and wire it in: create a project with the core in its files and one standing instruction, “read this before answering,” or reach it directly through a connector so the tools you already use can read it in place.

Then the habit that separates living from dead:

Here’s what changed since the last update: [ramble it]. Update the knowledge base — show me the edits first, then apply them and refresh the “Current as of” dates.

Staleness is how knowledge bases die, and the dates keep yours honest — a page that says “Current as of” three months ago is telling you the truth about itself.

---

Check your understanding: what’s the difference between the core and the shelf — and what’s the acceptance test for a knowledge base?

Use this process for other shared knowledge. Interview → hot core and cold shelf → stranger test → update ritual. That’s the client handbook, the team onboarding doc, the vacation handover, “how we do things here.” Every background explanation you’ve given twice is a knowledge base asking to exist.