A wiki your agents write and you read

Once you run several coding agents at a time, it gets hard to keep track of them. Each session makes decisions and raises questions, and nobody rereads the transcripts. CLAUDE.md and agent memory help the agent. They don’t show you what’s going on.

One fix I saw on Hacker News: agents keep a short wiki written for the human, with decisions, open questions and where each area stands. You read it, decide there, and the next session builds from the page you changed.

I’ve been doing a rougher version of this for about a month on static.md, the notes app I build. Here’s the setup.

What it looks like in practice

On Sat 24 Oct I had several agent sessions running on launch work. One prepared our ChatGPT app submission and another audited our listing in Claude’s connector directory. Each kept its own note.

The directory note opens with a verdict: live and healthy, two things stale, one item never done. Below that are the rules it checked, with sources, and the exact edits it proposed. I read the top and said yes to the edits, and the note’s last section records what was submitted the next day.

A second note started from a competitor check and became a list of feature ideas. I picked what to build, and the note split into Building now and Later. Each item under Building now went to its own agent and ended in a PR for me to merge. Later holds the rest, along with the questions still open on them. When I come back to those, they’re on the page, not in a chat I’ll never find again.

The setup

Connect your agent. For Claude Code:

claude mcp add --transport http static-md https://static.md/mcp

Then run /mcp, pick static-md and choose Authenticate. Claude, ChatGPT, Cursor and VS Code are covered on the agents page.

Then add this to your CLAUDE.md or AGENTS.md:

## Project wiki

Keep a short wiki for me in the static.md folder wiki/<project>.
Write it for me to read, not for you.

At the start of a session, read the note "<project> index" in
that folder, then the pages your task touches. If I changed a
page, follow my version.

When you finish a piece of work, update the pages it affects.
Record only what a new session would need to carry on from here:
- decisions: what we chose, who chose it, when, and what we
  rejected
- open questions, each with an owner
- a few lines on where this area stands

Keep each page to about one screen. Link pages with [[Title]].

Keep "<project> index" current: one line per page, then the
questions waiting on me.

Set each new page to "only me" right after creating it, and
tell me if that fails. Never put secrets in a page.

New notes start open to anyone with the link, so the instruction sets each page to “only me” right away. There’s a short gap between the two calls, so secrets stay out either way.

Here’s a small sample an agent wrote in this format, about how agents work with static.md. It’s public so you can read it.

Why notes instead of files

Files in your repo are a good default, and for one agent on one machine they may be all you need. Notes help in these cases:

• Several agents on one page. An agent in Claude Code and another in Cursor can write to the same note while you type in it. Edits merge, and if an agent rewrites lines someone changed since it read them, the edit is refused. I wrote up how that works and how I tested it.
• Links in both directions. Reading a note returns the pages it links to and the pages linking to it, so an agent changing a decision can see what depends on it.
• No local server. A page is a link, so you can read and edit it from any device, and share one page with a teammate without sharing the rest.
• A way to check the agent. Every agent edit stays in the note’s history for 30 days, with a diff and a restore.
• A fence around it. When you connect an agent you can limit it to one folder. Everything else looks like it doesn’t exist.

If you want a visual map on top, an agent can draw one as a whiteboard and link it from the index.

static.md also has a /memories folder for the agent’s own long-term memory. Keep the two apart: memory is what the agent needs, and the wiki is what you need.

Try it

Setup is on the agents page. If you try it, I’d like to hear which pages were useful and what went wrong: salut@static.md.