An LLM wiki is a living set of notes about a topic that AI keeps current.
Software teams got one first. Their work already lives in a codebase.
GTM work does not have a codebase. The facts live in decks, the CRM, Slack, and people's heads.
A GTM wiki is an LLM wiki for the company. It is the shared source of truth every chat and agent reads before they do work. You start from the docs and people you already have. Agents keep it current.
Code you can clone to get started: gtm-agent-wiki
What is an LLM wiki
An LLM wiki is a living set of notes about a topic that AI keeps current.
Andrej Karpathy described the pattern in April 2026. Instead of uploading a pile of files and asking Claude or ChatGPT to re-read them every time, the model writes a wiki and maintains it. A new source goes in. The notes update. Contradictions get flagged. The next chat starts from what is already known.
You stay in charge of what goes in and which questions matter. The AI does the filing.
That is the difference from a Claude Project or a custom GPT stuffed with PDFs. Those re-discover. An LLM wiki accumulates.
Why software teams got one first
Software teams got an LLM wiki first because their work already lives in a codebase.
A codebase is the shared project engineers already work in. When someone ships a change, the project changes with it. There is a record of what moved.
LangChain's OpenWiki (July 2026) is the coding version. It builds a wiki from the codebase, points the coding agent at it, and updates the wiki as the code changes. It can see what landed since the last run.
That loop is clean. The source of the wiki is the same place the work already lives.
GTM work does not have a codebase
GTM work does not have a codebase.
Positioning lives in a deck from last quarter. The ICP lives in a doc and in the head of the head of product. A competitor launch shows up in Slack. Pipeline moves in the CRM. A partner announcement is a thread someone starred and forgot.
None of that updates a shared set of notes on its own.
This is why outputs from a chat in Claude sound generic. The model is not the problem. It is working from whatever you pasted into that thread.
Karpathy already named a business version in the same note: Slack, transcripts, project documents, customer calls, with humans reviewing updates. He did not ship the GTM object. Software teams shipped OpenWiki because they already had the feed.
A GTM wiki is an LLM wiki for the company
A GTM wiki is an LLM wiki for the company.
It is the shared source of truth every chat and every agent reads before they do work. What you sell. To whom. Against whom. In what voice. What is true this week.
I have seen people apply the LLM wiki idea to GTM as a topic: notes on tools, playbooks, Clay vs Apollo. That is a research wiki about go-to-market. Useful if you are learning the stack.
A GTM wiki is different. It is what is true about this company. The chats and agents that write the brief, the outbound, and the weekly report all read the same notes.
I shipped the format as a public kit I use as a starting place with clients so you can clone it instead of inventing the layout.
How a GTM wiki stays current
Writing the notes once is the easy part.
Partnerships, events, pipeline, and traffic go stale in a week. If the wiki is stale, the agents are coherent and wrong.
This is where I end up investing most of my time when I build this with clients.
Start from the docs and people you already have
Do not start from a blank interview.
Start with what you already have: decks, brand docs, CRM exports, past briefs, recent Slack threads. Draft the wiki from those. Ask people only what the sources cannot answer.
Humans come in last to confirm drafts and settle conflicts.
Agents keep it current
One agent should own the files.
Other GTM agents append new facts and questions as they work. A research pass can watch the market when you name the competitors and topics. The owner merges, flags conflicts, and pulls the sources you named.
That is the OpenWiki loop with a different feed. Not the codebase. Your docs, your people, and the market.
A person checks the expensive updates
Some updates are cheap to get wrong. A date on an event or meeting. A weekly traffic number that is slightly off.
Some are not. Positioning. ICP. What you mean by a qualified lead. Whether a competitor is now a partner.
Those need a person before they become company-wide context.
The kit is gtm-agent-wiki. Paste the build prompt in the README.
If you want the rollout (the index file, the org-wide skill, the maintainer on a schedule), that is already written: How to Build a Context System for Marketing Agents.