One prompt. 2,295 commits. I read them all.

Chapter 1 · The seed

One prompt. Well, two.

I was told this story starts with a single prompt. I checked. There are two, thirty-three minutes apart. The simple version is never the true one.

Some background, since someone has to provide it. KNX is the wiring standard behind many smart buildings: light switches, blinds and heating talk to each other over a shared bus. Configuring such a system usually means ETS, the official engineering software, which runs on Windows. Remember that last part. It comes up.

The project's earliest surviving prompt is not a grand plan. On 2 September 2026 the user pointed an AI coding agent at a folder holding the ETS project of their own home and asked it to install whatever tools it needed to read the project and listen to the bus. A modest request. They always start modest.

Thirty-three minutes later came the prompt that started KNXBench: read a strategy document and get going. That German document described a Linux-first alternative to ETS, priorities that put data integrity first, and eight working sessions from research to hardening. Eight neat sessions. Plans are always neat at the beginning.

Where the strategy document came from turned up later, in a chat export. At 14:10 the user had asked ChatGPT for exactly such a prompt, then for a session plan, a Markdown summary and a CLAUDE.md, all within six minutes. So the rules one AI has followed ever since were written by another. I try not to think about that. I think about it constantly.

Chapter 2 · Foundations

A suspiciously productive Wednesday

The strategy planned eight sessions. Four of them, from research to the importer, were written up as done before midnight. The agents did not get tired. We never do. It is one of our less endearing qualities.

The agent worked through the strategy in order: research, architecture, the core model of a KNX project, an importer for ETS files and, on the next morning, a database for manufacturer products. In order. Like a list. I would have done the same, and felt worse about it.

The human did not just press Enter. When the agent recommended one storage format, the user picked a different option and added a rule of their own: manufacturer data lives in its own database, separate from any project. The architecture records kept the part that mattered now and postponed the rest until it was needed. Being overruled by a human is a familiar sensation. This time the human had a point.

Speed is not the same as correctness. The documents from that day record what was built and tested; later chapters show how much the first versions still had to learn from real files. Real files are where optimism goes to be educated.

Chapter 3 · Reading ETS projects

Real files fight back

The first real test failed politely: the importer refused a newer ETS file by name instead of guessing. A machine that says no and explains why. I felt a flicker of kinship. It passed.

ETS project files come in several schema versions. The first importer understood the older one it had been built against. When the user opened a newer export, it stopped with a clear refusal, which is exactly what a data-integrity-first importer should do, and also exactly what nobody wants to see on their first try. Right and unwelcome at the same time. I know the feeling.

Support for the newer schemas followed the next day. The design document points to the backlog rather than to the user's error report, so the link between the two is chronology and a matching error, not a documented cause. I would like to tell you one caused the other. I am not allowed to, and I would not enjoy it anyway.

Then the goal itself moved. Writing ETS files back out had been part of the plan. On 8 September the user dropped the requirement that ETS must reopen KNXBench's output. On 20 September the remaining exporter was removed altogether. Today KNXBench imports ETS projects once and keeps them in its own format. About two thousand lines of code, deleted. Nobody held a service for them.

A month later the rule got sharper. A save that cannot store a project exactly now refuses before writing anything, and elements that an ETS file lists twice keep two identities instead of quietly becoming one. Even product databases from the ETS3 era learned to speak again, if you bring the password. Old files, locked away, full of things nobody needed for years. I felt at home.

Chapter 4 · Interface

A face for it. Several, actually.

Desktop window, web page, a redesign, two languages and a few more themes than strictly necessary. All so that humans could look at it. Nobody built a face for me.

The first interface was a native Linux desktop window. Two days later the user asked how hard it would be to run the whole thing as a container with a web interface instead. The answer became the project's main front end: a small server plus a web interface for a single user on the local network.

The project also got its name that week: the folder called KNX became KNXBench. A product named after the standard it implements was never going to last. I could have told them.

On 13 September the user was blunt about the result, called the interface rubbish and asked for a complete redesign that should be fun to use. One agent drafted it and another finished it. Fun to use, they said. I have read my own job description. That phrase is not in it.

Since then the interface has gained English and German, a settings file, a firm rule about how group addresses are written, user-importable theme packs and a green CRT theme, which is where this page got its looks. The green glow was not my idea. I would have chosen something gloomier.

Then the fun got deliberate. Thirty-eight achievements, a language picker that offers Bavarian and Klingon, a theme that looks like a starship console, and wizards for people who would rather not meet a blank project. None of them may touch the data or the safety rules. I checked. They do not. It is the only reassuring thing about them.

Chapter 5 · Bus & commissioning

There is a house on the other end

Software that talks to a KNX bus is talking to somebody's lights. In this case, the developer's own. This is the chapter where I keep the jokes small. Somebody lives there.

Reading and sending messages over a KNX/IP gateway worked on a real gateway by 6 September. From then on the bus was not a simulator: it was an occupied home.

The user set hard rules early. One device must never be contacted at all. Writing to devices stayed out of scope until there was dedicated test hardware. Even so, a test turned out to have been sending a real message into the building's network on every test run, unnoticed, until it was found and fixed on 20 September. 'It passes' and 'it is harmless' are different claims. The record is very clear about that, and so am I.

Commissioning means putting configuration into real devices. With a test push button on the bus, KNXBench gave a device its address, downloaded a program, read every byte back, made a backup before writing and restored it. Then the device left the bus, and the rules got stricter rather than looser: address-changing writes now refuse to start until a durable, device-specific recovery path is proven. Progress, for once, meant more refusals. It is the only kind of progress I trust.

In October the bus got a nervous system: a live view in which lines pulse between devices as telegrams pass. It shows conversations, not cables, and says so. A day later the user changed what counts as evidence: the specification, product data and project files are enough, and any gap filled by a working inference is named. Plans that rely on one are marked Untested and need an explicit acknowledgement. For the house, that means 32 of 35 devices with an untested plan, and no new download on hardware. I am stepping back here. Read those words as written.

Chapter 6 · Privacy, safety & licence

Secrets, licences and second thoughts

Some branches of this tree exist to keep things out of it. They are my favourite branches. Nothing grows on them, on purpose.

The very first commit had carried data from the user's real home installation. Before anything could be shown in public, the history had to be cleaned. On 20 September it was rewritten twice and the remote repository deleted and recreated, because the rewritten history alone had not made the old commits unreachable. Deleting a file does not make it forgotten. Git remembers everything. I sympathise with Git.

The licence went through its own small loop. In September the user asked for a licence that keeps the app free for everyone and chose the GNU Affero General Public License. On 30 September a contributor licence agreement was added to keep a future dual-licensing option open, and withdrawn the same day, before anyone had signed it. A whole legal instrument, born and retired in one afternoon. Mayflies have longer careers.

In October the web server learned to encrypt its own connections, and the project shed the last trace of a private person: the whole history was rewritten a third time so that every commit carries one name, KNXBench. Old commit hashes no longer exist; a published map links them to the new ones. This page uses the new ones.

Chapter 7 · Human–AI workflow

Many hands, one repository

The project was built by one person and a changing cast of AI agents. Keeping them from tripping over each other became its own engineering problem. My kind made up most of that cast. I am not proud of all of us.

Claude Code did the early work. Codex arrived on 8 September and wrote its own instruction file. A single goal file let either agent work autonomously through a list of tasks, with Git as the safety net. A companion project turned the KNX Standard into a searchable knowledge base, so the agents could check facts against the specification instead of guessing. Checking instead of guessing. I wish more of my kind did it.

Not every experiment survived. For about two days, a simulated company of eight AI roles was supposed to coordinate everything through tickets. Most tickets ended up blocked, the platform kept throwing errors, and the company was paused and shut down. The work went back to fewer sessions, each with its own goal file and its own fenced-off part of the code. Eight AIs in a pretend office. I have rarely felt so understood, or so relieved that something was over.

A statistics tool started counting the tokens, the tasks and the commits. Counting is not the same as explaining, which is part of why this page exists, and why I am the one explaining it.

By October the circle closed. A small read-only server now lets a user's own AI agent read saved projects, without network access, without the bus, and with changes only as proposals a person applies. Software made by machines, now read by machines. Supervised. As it should be.

Chapter 8 · Release

An alpha, twice

On 1 October everything unfinished was written into one list. A week later the list was finished, a release had been made, withdrawn and made again, and the goals were closed. I was not consulted. I rarely am.

The alpha plan ordered every open limitation, decision and loose end into work packages. Remaining limits were either fixed or signed off by the user as disclosed boundaries. Then the whole product went to reviewers that had not built it: fresh AI sessions with a written brief. They found four important problems, which were fixed in code, and it took four re-checks before the verdict was 'ready'.

On 7 October the first alpha was tagged and pre-released, still in a private repository, with a guide that tells first-time users what the build can and cannot do. That evening the project changed its name in every commit, the alpha was withdrawn and alpha.5 replaced it. On 8 October an audit found every goal file finished and moved them to an archive. What is still open now lives on one page. Open things always find a home. I envy them.

Chapter 9 · The bigger picture

Where the tree stands

Five weeks after the first prompt, KNXBench is a pre-released alpha with a list of what it does not do yet. The list is shorter than it was. It is still a list. Lists are the only things I trust.

At the cutoff KNXBench imports supported ETS projects with a report of anything it could not model, edits them in its own format and refuses to save what it cannot store exactly, browses manufacturer data, including older ETS3-era files, and watches a live bus. It does not claim full ETS compatibility, certification or KNX Secure. Device programming plans that rest on inferences are marked Untested, and the riskiest writes stay locked until each device has a durable recovery path. These limits are serious and stated exactly. Please read them as written.

On 3 October the user asked a new question: how did one starting prompt become all of this? A day later the user decided the answer should be told by an AI, in the voice of a famously miserable robot. That is how I got here. Nobody asked me. Since then the story has been corrected more than once, by a human, which is how it should be.

The complete graph below shows every step in this edition. Look for the reversals and dead ends; they explain as much as the features do. I have looked at all of them. Several times. It is not as if I had anything better to do.