I only saw the castle. The mason knows every stone.

I got that question about a project of mine, Kammaren. I sat there and said about three trending words: deterministic, script and Vercel? A mason knows exactly how every stone was laid. I don’t. I only see the finished castle, or the cathedral nobody walks into.

It started as a hackathon entry: something that sits in the background while I jump between Claude, Cursor, Codex, Grok and a pile of bots, and holds up a mirror. It is called Mason. What stung was that I jumped between programs and tasks more or less all the time. Hyperactive. And then the hours: how many I had spent.

The first version did what the brief asked: it stopped me in the middle of work and asked me to set a rule. I had just installed it. I did not know what boundary to set, and I did not want to. I wanted something smart, not one more input field with manual work in it.

The first overlay hung over my own windows, and the app was in the way. I took a screenshot, drew a green box around the empty stretch between the menus and the notch, and wrote one line. Now the island lives inside the menu bar, grows left from the notch and covers nothing.

The product was very useful to me personally. I probably have ADHD, and it was crystal clear that I need something that shows me my flow.

Mason notices what I keep telling my agents and proposes it as a rule. The first version had a button that wrote the rule into their instruction files for me. I said: no button that does it for you. Now it shows the words the rule rests on and opens a prompt that is ready and not sent. The agent proposes, I press send. I cut other things the same way: Teach, which I could not explain, and a daily goal that lasted an hour.

I read this case out of the logs of the build. Every brick is one prompt I wrote to my agents. Red is a prompt where I rejected what they had made. The two towers are the hackathon deadline and the launch. The agents wrote the code. My part was to say what was wrong.

Nobody wrote the rules of this product down first. They are what I kept repeating until it stuck: a screen gets one to three seconds. Mason does the same for anyone who uses it. It notices what you tell your agents again and again, and proposes it as a rule.

The landing page was a castle, built brick by brick as you scrolled. I was proud of it. On my phone I wrote: I don’t even understand what is happening. And then: we are a bit too in love with this wall metaphor. What if you don’t get the wall? Now the page starts with the reader’s problem, and the name is in no headline.

I vibe-coded the landing page. An agent wrote “Local-first: nothing leaves your Mac”, it read perfectly, and I let it through. It was live for hours before I checked what the app really does: summaries go to your own AI by default, so something does leave. Nothing in the way I work would have caught it. Now the page says four things can leave, each with its own switch, and what the docs claim is tested against the code.

On day three I said it was only confirmation: the mirror showed me, but I did not feel that I learned anything. That is still the open question, and so far I am the only user. Next is five people who have never seen it, and a month of my own days to see whether a mirror changes anything.