Blog
Thoughts, methods, field notes. Newest first. Also available as RSS.
- One lead, many workers: moving a compliance system into an ERPI froze my own regulatory app and rebuilt it as modules inside an open-source ERP, with a lead agent, workers and reviewers. How two million tokens of source became a spec, what the old design could not tell apart, and the bug every review found: the agent could lift its own gates.
- A phone number for the agentsA phone line that calls me with a status, answers when I call, and takes a message from everyone else. The pipeline on one gaming graphics card, the latency budget, two telephony traps, and the rule that a stranger never reaches the model.
- Nineteen of my agent's skills had gone silentAn audit of what Claude Code really loads at the start of every session: a skill menu with a shared budget that drops descriptions without telling you, a memory index that stops at 200 lines, and a second memory store I did not know I had.
- 233 treasure legends, 126 newspaper reports, one mapA research desk for lost things in Upper Austria: open find registers, old newspaper reports and legends on one map, and what was left of the best story after its first print and a parish register were read. Desk work only, and the law says why.
- Eight models read a letter from 1726One page of an 18th-century Spanish secretary hand, the same prompt to eight vision models. The error rate matters less than how each one fails: some slip on a letter, some swap in a plausible wrong word, and two write fluent fiction.
- The dead drive was just Mac-formattedA 2 TB external disk that Windows had shown without a drive letter for years. It was healthy the whole time. How to tell from the partition type, how to read it for free, and why its big ZIP files look corrupt when they are not.
- A usage meter for Claude Code that stays on screenAn always-on-top Windows widget that shows your Claude subscription limits, plus an F9 key that brings the Claude terminal up and puts it away. Plus the three things that broke on day one, and an update with a daily budget and token statistics.
- An agent team that ships Amazon listingsResearcher, marketer, copywriter, designer, publisher: how a pipeline of specialised AI agents takes a product from a folder of photos to a live listing, and where the humans stay in the loop.
- An agent that argues both sides: legal research with RISHow a landlord who wanted to keep our deposit taught me the method, and how it became a system of agents that fetch Austrian law and case law, argue for and against, and refuse to cite anything they have not read.
- CAD for agents: how a scaled plan gets a plot onto the roadA building plot with no frontage needs a legal connection to a district road. The deciding document is a plan the road authority can measure. Here is how an agent draws one, and why every number on it is re-measured.
- RapLab: never let the model count syllablesA lyric workbench where the language model writes too much on purpose and a small deterministic judge decides what fits. Plus four prompt traps that cost real hours.
- Run your agents from your phone with TailscaleA private network between your phone and your workstation, SSH with keys, and a session that survives a dead zone on the train. No open ports, no cloud VM. Plus the one bug that ate an evening.
- Hello, worldBoot message. What this site is, and what will end up here.
- How this site worksA terminal that is really a static website: the stack, the themes, the command line and the game.