Best practices for an MCP wiki

Connecting an AI to a knowledge base is the easy part. Keeping that knowledge base worth reading is the work.

An MCP wiki is a set of notes your AI reads and writes over MCP. Claude and ChatGPT fetch what they need and write new notes back. Get the structure right and it compounds. Get it wrong and it rots like every wiki before it.

Here are eight rules that keep it useful.

1. Small notes beat big documents

One note, one topic. A model can fetch and update a short note cleanly. A 4,000-word page means it rewrites the whole thing to change one line, or worse, it skips the update.

If a note covers three subjects, split it into three. Short notes are easier to link, easier to find, and easier for your AI to maintain.

2. Folders and tags that match how you think

Use notes, folders, and tags. Folders for where a thing lives. Tags for the themes that cut across folders.

Name them the way you would say them out loud. Fixes, Decisions, People. Not a taxonomy you will have forgotten in a month. The structure is for you first and the AI second.

3. Write instructions per folder

This is the rule most people miss. Tell your AI how to maintain each part of the wiki.

A Decisions folder can carry instructions like "every note starts with the decision in one sentence, then the context, then what we ruled out." Now Claude and ChatGPT write notes that fit your system instead of inventing a new format each time.

4. Always write a summary

Every note gets a summary. Summaries are what search and recall read first, so a good one is the difference between a note your AI finds and a note it walks past.

One or two sentences. What the note is, why it matters. Let the AI draft it, then keep it honest.

5. Link related notes

Connect notes with wiki-style links. A decision links to the project. A fix links to the error. A person links to the meeting.

Links are how context compounds. When your AI reads one note, the related ones are one hop away, and synthesis happens at write time instead of being re-derived on every question.

6. Let the AI do the bookkeeping

The reason wikis die is maintenance. Humans get bored of tagging, linking, and deduplicating. Models do not.

So hand it over. Ask your AI to file the note, tag it, link it, and flag anything it contradicts. The loop is simple: research in your AI, save to the wiki, act, repeat.

7. Search before you create

Tell your AI to search the wiki before adding a note, and to update the existing one when it finds a match. Otherwise you get five near-duplicate notes about the same thing and a wiki nobody trusts.

One fact, one home. Update it in place.

8. Keep it client-agnostic

Do not wire your wiki to a single tool. The whole advantage of MCP is that the same notes are read by Claude, by ChatGPT, by Cursor, by whatever you use next.

Capture a thought in one client, refine it in another, query it from your phone. If the wiki only works in one app, you have rebuilt the local-vault problem with extra steps.

The short version

Small notes. Folders and tags you would actually say out loud. Instructions per folder. A summary on everything. Links between related notes. Let the AI keep house. Search before you create. Stay client-agnostic.

Do that and the wiki earns its keep.

---

Ready to build one? The LLM wiki build guide walks through the setup, and the Knowledge Wiki template gives you the structure and per-folder instructions out of the box.

For the thinking behind the pattern, read why Karpathy's LLM wiki is right. If you are not sure you need more than a single Markdown file yet, LLM wiki vs. plain Markdown shows where the simple version stops working.