The first twenty minutes

The do-it-yourself path needs no onboarding call or sales conversation. Four steps, and the last one is a single keystroke. All four work today.

1 · Make a board — about five minutes

The board is where you put the work and where the operator reports back. It runs on TaskPeace, which we build and run ourselves. Sign in with Google and you have a board; it is free while TaskPeace is in beta, and there is nothing to configure.

Then write down five or six things you have been meaning to get to. Plain sentences, the way you'd say them to a person: “the pricing page is embarrassing”, “nobody has looked at whether search traffic is down”, “the contact form might be broken”. You are not writing specifications. Ordering matters more than wording, and you can reorder any time.

If you leave the board thin, that is fine too — when it runs out of your work it goes looking for what should have been on the list, and files what it finds.

Not ready to sign in? The same board runs in local-only mode, in your browser, with no account at all. You can look before you commit to anything — though the operator needs the signed-in version to reach it.

2 · Install — about two minutes

On the board, open Account and click Install MCP. It hands you a single line with your own access token already in it. Paste that line into your terminal on the Mac you want the operator to work from.

PROMPTPRIO_API_TOKEN=pp_… bash <(curl -fsSL https://taskpeace.com/install.sh)

That line checks your token, installs the small connector that lets your agent read and write the board, and adds a one-key resume so a bare q picks up a stopped run. It asks for nothing else — no hosting passwords, no repository access — because it works through the tools you are already signed into on that machine.

You can also run curl -fsSL https://taskpeace.com/install.sh | bash and it will ask for the token; copying the pre-filled line from Account is simply fewer steps. Either way, read the script in a browser first if you would rather. We would. And never paste that token anywhere but your own terminal.

3 · Restart your agent — ten seconds

Quit Claude Code and open it again. It only looks for new connections at startup, so this step is not optional — skip it and the next one does nothing. To check it took, ask it to list your tasks; it should read the board you just made.

4 · The first run — two to ten minutes to a plan

The very first time, give it something to read. Paste this into your agent, from inside the folder of the business you want worked on:

https://leftrunning.com/first-run.txt

Timed on two businesses it had never seen (2026-09-06): about two minutes for a four-file folder with no code in it, about ten for a site with 1,300 pages. It reads everything you have before it writes anything, which is why the range is wide. Working the first item then takes as long as that item takes.

It reads what you actually have — the repository, the live site, what you sell — writes a short plan onto your board in your own words, does the first item on it, and shows you the proof. It asks you at most one question, and only if there is genuinely nothing to read.

One thing that will look confusing: when the installer finishes it suggests a different prompt, taskpeace.com/autopilot-prompt.txt. That one is the general TaskPeace autopilot, written against our own portfolio and the tools it already has wired. Use the link above for your first run instead — it is the same operator, pointed at a business it has never seen before. After the first run neither matters: you press a key.

After that first run you do not paste anything again. d does exactly one thing, shows the proof, and stops — do that two or three times, until you have a feel for the quality of its judgement. Then, on a night when you have something better to do, press p and go to bed.

When it goes wrong

• Nothing happens when you press a key. The restart did not take. Quit the agent properly and reopen it, then ask it to list your tasks before pressing anything.
• It asks you for something. That is O working as intended — it hit money or your name, and it is handing that back. Everything else it does alone.
• It picked something you did not want. Say so, in a sentence. It re-picks. You do not have to write a prompt to steer it.
• It shipped something you dislike. Everything is an ordinary commit and an ordinary deploy, so it comes back out the ordinary way.
• It stopped. It reports why in one line — and if it was paced down because the night was long, it comes back when the budget resets.

Want the first run installed with you?