One sentence became my lamp's rule

On 20 October 2026 I wrote this into a conversation, about the evening light of the lamp in my room:

“it should by no means start any whiter than what was on before” Kristóf Karner, 20 October 2026 — translated from Hungarian

The same day it became a rule in the lamp's controller. In the evening the light only moves towards warmer, and not for a moment does it get whiter than what was on before. Late in the evening all three bulbs burn at their warmest white, and only the brightness goes down. I built the lamp in a conversation with Claude, in a single day, in four versions.

Other sentences turned into other systems:

• “it should only let me out of the recap once I've paged through it” — became a rule of my e-reader's firmware: from the next day's summary of what I read, the only way forward is to page through it.
• “I'm forced into constant hunger” — I said this about a streaming service's recommendations. It turned into an offline music player for my own collection, with no recommendations in it.
• “it shouldn't have to be exactly twice up and down” — the wall switch became a signal: if I flick it up and down a few times quickly, the lamp switches to bright light for cleaning.

None of them is a hard program. In 2026 an AI agent works on its own for hours on a task, and it can build into an existing open-source tool too. So the system comes down to whether I can say how I live, and whether I can check that it works that way.

What stayed my job

I built them with AI. I decided what they should do, and how I check them. In practice, that was three things.

Putting it into words. My requests were mostly rules about how I want to live: in the evening the light shouldn't get whiter, the player shouldn't recommend, the reader shouldn't let me go on until I've paged through what I read last time.

Saying no. The AI often suggested something else. For the e-reader it recommended a dark, cinematic main menu; I chose another. Once it suggested a week's break so the time could go to my business; two more new versions were built that day.

Trying it. Most of the faults only showed up on the real device: that the type doesn't fit the screen, that the wake-up doesn't work on battery, that the lamp reacts to flicking the switch differently than I expected. Only I could check that, and it was the larger part of the work.

What it cost

Time, attention and quota. Between 31 May and 21 October 2026, AI worked on my systems on 71 days. In Claude Code that's about 121 hours of machine time in total, about 45 of them on the e-reader, across 14 days. The lamp was a single working day.

In money, little. The e-reader's cloud bridge runs on a free plan, and my subtitle translator made the Hungarian subtitles of a film with 972 lines in 4.4 minutes, for 4% of the five-hour Claude quota (measured on 20 October 2026). The quota has a floor, though: on 21 October 2026 the planning of this very series stopped on it, because six AI helpers running in parallel used it up at once.

Where it breaks

If I don't understand what it built. If the person directing doesn't understand what the AI built, every fix becomes another request. Geoffrey Litt writes that understanding is now the bottleneck. That's why I keep notes for every project, which both the AI and I can read back to see what we decided and why.

If I connect everything. My e-reader's bridge sees private data and outside text at the same time, and it also talks outwards. Simon Willison calls these three together the lethal trifecta. That's why the bridge has a daily call cap, and why the text I read is deleted after processing.

If the provider changes something. One person and one AI provider: if the terms or the quotas change, the cost of maintenance changes too. That's why my systems are made of simple files, with few dependencies, and the code is on my machine.

Open source became malleable

The research group Ink & Switch argues in a 2025 essay that open source gives access to the code, but changing it takes expertise, and AI mostly produces more stand-alone apps. Most of my systems build into an existing open tool: the e-reader into my own branch of the CrossPoint firmware, films into the IINA player and mpv, the lamp into the bulbs' local protocol. The agent brings the expertise, the existing tool provides the place.

Conversation becomes configuration

My systems are configured in text: the project's notes (the CLAUDE.md), the agent's memory and the odd skill. Whatever I've said once — that there should be no artificial line breaks, that I want pronunciation in Hungarian letters — goes into the notes, and the next session starts with it. Since September 2026, Notion can also hand a note page to Claude Code as a skill.

I could show the code, but not the configuration, because it holds my schedule, my taste and my limits. That's why every project of mine lives in a local git repository, with no remote.

The connections

The systems touch each other. My Notion calendar turns into an e-reader book, and what I've read gets ticked off. When a film starts in IINA, the lamp dims. The keyboard's play key goes to whichever player is open.

My favourite connection runs between two Claudes. On 29 September 2026 a chat version of Claude used a skill to write the accompaniment for a silent film, and put it on the memory card as a playlist. The same day, my music player's Claude Code session called the same skill and created the list in the player too. Between the two Claudes, my clipboard was the bridge.

The whole map, with dates and a glossary, is on the Systems of my own page, and the e-reader's details are in a separate piece.

What else is possible

The e-book built from a calendar works for anyone who plans their reading in a calendar. The subtitle translator would work for any language pair, small languages and old films included. The structure of the recommendation-free player (an SQLite database, Python's own libraries, plain JavaScript and tests) also suits the finite collection of a label or a DJ archive. I build systems for others the same way; my piece on automated systems is about that.

Questions on the topic

← Back to Writing