Five prompts from an empty account to a working AI agent

“Set up in one conversation” is the kind of claim that only means something if you can see the conversation. So here it is: the prompts, in order, and what your coding agent does with each. Run them in Claude Code, Cursor, Codex or any editor with Verb connected (one command, here), inside the project you want the AI agent in.

1. Create the site

Create a Verb site for this app.

Your agent reads your README and landing copy to write a short description of your product, works out from the code whether your users sign in, tells you what it used, and creates the site. You get a live site and a test site beside it. The test site only answers on localhost, so everything that follows happens on your own computer until you decide otherwise.

2. Decide what the AI agent may do

Write a verb.md for the things our users most often do, show it to me, then check it with Verb before importing anything.

This is the step that matters. Your agent maps your app, picks the actions a user would actually ask for rather than every route you have, and writes them into one file, each marked as reading, writing or deleting. It shows you the file and why each action is classed the way it is. Then it runs Verb’s own check over it without saving anything, so a problem shows up now rather than after you import.

Read it. It is the list of what an AI will be allowed to do inside your product, and it lives in your repository like any other file, so it changes by pull request from here on.

3. Put it on the test site

Import it into the test site.

Actions that only read are on straight away. Actions that change data arrive switched off, and you switch on the ones you want, by asking.

4. Add the line to your app

Give me the script tag for the test site and add it to the root layout.

One line, in your layout rather than a single page. Start your app locally and an Ask AI button appears. Ask it something your users would ask. It answers from your real development data, as whoever is signed in.

5. Let actions run for real on your computer

Turn on real actions for the test site.

Until now, anything that would change data was simulated. This is the first of the two moments your agent stops and asks: it tells you that actions will now really run against whatever your local app is connected to, and waits for your yes. Every change a user asks for still shows them the real details and asks them too.

When you are ready for real users

Add https://app.yourproduct.com as the live domain.

Your agent adds it and gives you three ways to prove it is yours: a meta tag, a DNS record or a file. Once the widget has loaded there, confirming the domain is the second moment it stops and asks, because that is when your users can reach it.

How long it takes

With actions that only read, the five prompts above are minutes. Actions that change data take longer, and they should: they run against your own backend, and the right amount of testing on your own computer is more than one question. What is gone is the part that was never the point: the tabs, the keys and the copying between them.

Building and testing is free, with no card. Connect your editor and paste the first prompt.