.

Most of my energy these past few years has gone into one thing: collapsing frontier AI capability into products an ordinary person can just open and use. The models change every week; what survives is whatever actually solved a problem.

Hence Vibe Coding — ship v1 within the hour, then let real feedback decide where it grows. The code I delete usually matters more than the code I write.

Three terminals stay open while I work: Cursor for anything with a screen attached, Claude Code for refactors and migrations, Codex for the long jobs I would rather not sit and watch. Whichever fits the moment gets used, with no stack to defend. What is scarce was never tooling — it is judgement.

cn2an was the first of them, a small utility turning Chinese numerals into Arabic ones, now downloaded a hundred thousand times a month. Its lesson was plain: doing one small thing well enough that others depend on it beats ten half-finished ones.

Open source is part of that method. Publishing the code is the most honest demonstration of capability, and the shortest path to putting a tool in the hands of whoever needs it.

More of them died than shipped. The half-finished ones are all parked under an org called ai-graveyard — the name is a joke at my own expense, and a reminder that most ideas never reach anyone. The code in the graveyard stays public; anyone who wants to pick it up can.

Building alone, writing code turns out not to be the bulk of it. The rest goes to answering issues, rewriting copy, watching errors come in, and working out what the next version should drop. None of that hands off to a model yet.