The Team Knowledge Base for Claude and ChatGPT

Someone on your team has already built an LLM wiki. A folder of Markdown notes their AI reads when it needs context, so they stopped re-explaining the codebase every session. It works. It also lives in one person's vault, and nobody else can reach it.

That's the ceiling of the single-player LLM wiki. One person writes it down, one person's Claude or ChatGPT reads it. The rest of the team still starts from zero.

A team knowledge base for Claude and ChatGPT is the same idea, shared. One place your whole team writes decisions, runbooks, and conventions. A built-in MCP server so every teammate's assistant can read the same notes over the same protocol. Write once, and every assistant can reach it.

Why the personal version doesn't scale

The personal LLM wiki solves a personal problem: your AI forgets everything between sessions, so you save context it can reread. Give that to a team and three things break.

It lives in the wrong place. A local vault or a personal account is readable by exactly one person. Your teammates' AIs can't see it.

It goes stale in one head. The person who wrote it moves on, changes their mind, or forgets to update it. Nobody else can.

It gets re-explained anyway. The moment a second person needs the same context, someone pastes it into a chat. You're back to the tax you were trying to remove, just spread across more people.

A shared knowledge base fixes the location. Everyone reads and writes the same notes, and everyone's assistant reads them too.

What goes in it

The same things a good team already half-writes, in one place instead of five.

• Decisions. Why you chose the stack you chose, what you tried and rejected, the reasoning behind past calls. So nobody relitigates them.
• Runbooks and process. On-call, releases, incident response, the steps everyone is supposed to follow.
• Conventions. How you name things, how you review, the house style your AI should match when it drafts.
• Ownership. Who owns which service, who to reach when it breaks.

Put each in a folder with custom AI instructions baked in, so an assistant working in that folder follows the voice and the rules without being told each time.

But nobody reads the team wiki

They don't. That's the usual fate of a team wiki: written once, skipped forever, stale within a month.

This is different for one reason. Your team's AI can read it. When an engineer asks their assistant how deploys work or why you moved off Redis, it can answer from the notes instead of interrupting a senior engineer. Keeping the notes current pays off the moment someone asks their AI instead of a person. The wiki stops being a chore nobody benefits from and becomes something the team's assistants actually consult.

It also works across assistants. The same notes serve Claude and ChatGPT, so it doesn't matter which one each teammate uses. One base, every assistant.

Setting one up

The mechanics are boring, which is the point.

1. Create a team and a few shared folders: decisions, runbooks, conventions. The set up your team guide walks through it.
2. Add team AI instructions per folder so an assistant working in them follows the same rules.
3. Each teammate connects their own Claude or ChatGPT over MCP. Personal notes stay personal; per-folder roles decide who can edit what.

A team is free for its first 25 team notes, then it's per seat. The team billing page has the exact mechanics.

The point

The single-player LLM wiki proved the pattern: an AI that reads your written context beats an AI that starts from scratch. The only thing wrong with it on a team is that it's single-player.

Put your decisions, runbooks, and conventions somewhere every teammate's AI can read. See Hjarni for teams, or start from the knowledge wiki template.

Write once. You all remember.