How to build a company AI brain with AgentDocs

Credit where it's due: the four-part "company AI brain" framework is Ben Sufiani (The Captain) at Pirate Skills — read the original, The Company AI Brain. We loved the idea and wanted to show one concrete way to build it. This is our "here's how we'd do it in AgentDocs" take, not a replacement for his piece.

The problem: you keep re-explaining yourself

If you work through an AI agent, you know the tax. Every new conversation starts from zero. You re-paste who your customers are, how you write, what you decided last week, what's actually in flight right now. The agent is capable, but it has no memory of your world, so you spend the first ten minutes rebuilding context before any real work happens.

Ben Sufiani at Pirate Skills framed the fix well in The Company AI Brain: give the agent a memory system it can read on startup and write to as it works. He describes it as four markdown parts, and the shape is simple enough to keep in your head:

• Docs — the rules and knowledge. Who you are, how you work, what's true.
• Skills — repeatable procedures. The step-by-step for things you do again and again.
• Journal — an append-only log of actions and decisions. What happened, when, and why.
• Dashboard — the live state. Current priorities and what's in progress right now.

The loop is: the agent reads the current state, does the work, and logs the result. Do that consistently and the context compounds instead of evaporating at the end of each chat.

Why AgentDocs is a natural home for it

The framework is markdown and plain language by design, and that's exactly what AgentDocs is: a markdown-native, agent-first workspace — Docs, Sheets, Slides, Database, Drive and Pages — that your AI agent reads and writes over one MCP connector. You don't bolt a memory system onto a tool built for humans and hope the agent can drive it. The agent is a first-class editor of every surface, so the brain lives where the agent already works.

Here's how each of the four parts maps onto a surface.

Docs → Docs and Pages

Your rules and knowledge go in plain-language Docs. One doc for your positioning and tone, one for your customers, one for the decisions you don't want to relitigate. Because it's Markdown, it reads cleanly for you and for the agent, and it diffs like code when it changes.

When the knowledge gets bigger than a single page, Pages gives you nested pages and linked databases — a tidy place for a handbook that grows, without turning into a folder of loose files.

Skills → a Doc per procedure

A skill is just a written procedure the agent can follow: "how we publish a blog post", "how we qualify an inbound lead", "how we close the books each month". Keep each one as its own Doc, in steps, in the order you'd actually do them. The value is consistency — the agent runs the procedure the same way every time, and when you improve the procedure, you edit one Doc rather than re-explaining it in every conversation.

Journal → an append-only Doc

The Journal is the part people skip and then regret. Keep a single Doc where the agent appends a short entry every time it does something that matters — a decision made, a task shipped, a number that moved. New entries at the top, old entries left untouched. Later, you can ask "what did we decide about pricing, and when?" and the agent has a real record to read back, not a guess. If you'd rather have it structured, a Database with one row per entry works too — date, area, what happened — and is easy to filter.

Dashboard → a Database on a Pages view

The Dashboard is the answer to "what's going on right now?". A Database is the right home for it: a real structured table with your active projects, their status, the owner and the next step. Surface it on a Pages view and it reads like a status board. The agent updates it as work moves, so on startup it can open one place, see the current priorities, and pick up where things stand — instead of asking you.

How to start — small, then let it grow

Ben's advice is to start with one folder for one business area, and it holds up. You don't build the whole brain on day one. A sensible first pass:

1. Connect AgentDocs to your agent. It works in Claude and ChatGPT over a single link, with no install and no card.
2. Pick one area. Marketing, support, ops — whichever you re-explain most often.
3. Write the core rules in plain language. One Doc. What's true, how you work, what not to do.
4. Wire up one real workflow. Turn the thing you do every week into a Skill doc and run it.
5. Add the Journal and Dashboard once there's real work to log. Then let the agent keep both current.

Review it weekly. When the agent gets something wrong, it's usually because a rule was missing or stale — so fix the Doc, and it stays fixed. That's the compounding: every correction becomes permanent memory instead of a note that's gone by the next conversation.

A starter prompt

Once AgentDocs is connected, paste this to scaffold the four parts. Edit the bracketed bits to match your business:

Set up a company AI brain for me in AgentDocs, using four parts.

1. A Doc called "Rules" — our core knowledge: who we are,
   who our customers are, how we write, and decisions not to relitigate.
   Start it with a few headings and leave me prompts to fill in.

2. A Doc called "Skills" — one section per repeatable procedure,
   written as numbered steps. Seed it with one skill: [e.g. "Publish a blog post"].

3. A Doc called "Journal" — append-only. Newest entries at the top,
   each one dated. Never edit or delete past entries.
   Add an entry every time you do something that matters.

4. A Database called "Dashboard" with columns:
   Project, Status, Owner, Next step, Next step due (date).
   This is our live state — keep it current as work moves.

On every startup: read Rules and the Dashboard first, then ask me
what we're working on. When you finish something, log it in the Journal
and update the Dashboard.

That's the whole shape. Four documents, one connector, and an agent that reads the state before it acts and writes it after — so tomorrow's conversation starts where today's left off.

The four-part company AI brain framework is Ben Sufiani's (The Captain), at Pirate Skills — read the original here. This post is our take on building it with AgentDocs.