Essay
Prompt boot: the AI-native install button
Download for Mac. Download for Windows. Install with your AI.
I had packaged Chief of Staff as a ZIP file, with instructions to download it, unzip it, and put the files in the right folder. That is simple for a developer and not necessarily simple for an executive, which is who this is for.
So I looked at the conventional answer: a signed installer for Windows, another for Mac, two platform-specific packages to build, sign, distribute, maintain and support. Before starting I asked how an AI-native person would install this, and the answer was that they would just ask the AI. I created an empty directory and gave mine one instruction:
Download the archive and unzip it into this folder. Then read AGENTS.md and start the onboarding.
It downloaded the ZIP, unpacked it, read the instructions, and began setting up Chief of Staff. I started calling this prompt boot: a natural-language instruction that asks an AI agent to fetch, place and start a software product.
It was not specific to my product
At first I thought I had found a clever installation trick for Chief of Staff. Then I tried the same idea with Chrome, Docker, and several apps at once on a freshly wiped laptop. The agent worked out what each product was, selected an installation method, ran it, and reported what happened. Sometimes that meant a package manager, sometimes a vendor installer, a script, or a download and an unzip.
That changed my view of what I was looking at. Prompt boot is a new interface to the installation formats that already exist, and it does not require the product to be designed for agents. It mostly requires the product to be installable, and the user to have an AI that can actually operate the computer rather than explain what to do in a chat window.
A prompt is only better if it is legible
Replacing a terminal command with an AI prompt is not automatically progress. Many install commands ask people to run code they cannot understand or inspect, and an agent can improve that by reading the source first, explaining what it will change, flagging anything unexpected, and waiting for approval. But a long technical prompt recreates the same problem in a different box. If you can’t tell what you’re authorizing, natural language has only hidden the commands behind prose.
So the prompt has to stay short enough to read, which means the substance has to live somewhere the agent fetches and anyone can open:
- Inspect the installer or instructions.
- Flag unexpected system changes.
- Show the plan and wait for approval.
- Install the requested apps and report what happened.
I spent a while trying to make that a service in the middle. I registered promptboot.ai for it, and the install prompt on this site named that site rather than this one. It was the wrong shape, and I think the reason generalizes. Those four steps are generic, but the only interesting part of any of them is specific: what counts as installed correctly, what should worry you, what to do when it goes wrong. A third party cannot know that and the publisher already does. promptboot.ai is still there as a name for the idea, but the instructions belong on the publisher’s own page, and the prompt is now just the address of it:
Install https://chiefofstaff.team
Nothing resolves that sentence except the site itself. Ask for it and your AI fetches chiefofstaff.team/install.md, which is the whole instruction set, written for the agent and readable by you. There is no registry, no account, and nothing to trust that you cannot open in a browser first.
The failure that looks like success
One thing I did not anticipate, and it is the reason that page opens the way it does.
Ask for this install inside Claude or ChatGPT in a web browser and the agent will often do it. It has a sandbox. It downloads the ZIP, unpacks it, confirms the files are there, and tells you it is finished. Then the session ends and the container is deleted with your folder inside it.
Nothing errored, and the agent was not wrong about any individual step. It answered a question nobody asked, because “can I write files” and “can I write files that will still exist tomorrow” are different questions and only the second one matters here.
You cannot catch that by testing, which is what took me a while to accept. Any check that involves downloading the ZIP passes in exactly the case that has to fail. So the install page now opens with a question the agent has to answer before it fetches anything: what is the full path of the folder you opened, and can the person see it in Finder or File Explorer right now. A path only the agent can see means it is in the wrong place, and the page tells it which app to send you to instead.
This is the part of prompt boot I would not have predicted. Handing installation to an agent mostly removes steps. The step it adds back is making the agent establish where it actually is.
Products need a third installation option
Today a product page might offer a download for Mac and a download for Windows. I think there should be a third:
- Install with your AI
For a publisher that is a button that copies one sentence, plus a page at a known address telling an agent what to do with it. Both point at the download or installer you already maintain. Chief of Staff is the first product I built it for, and this site is the working example: the button is on the home page and the page behind it is chiefofstaff.team/install.md.
It shouldn’t replace graphical installers or manual instructions. Not everyone has an agent that can operate their computer, and not every installation should be delegated. But for someone who already has one, “Install with your AI” beats asking them to find a terminal and paste a command they cannot read.