How I work

With AI, building goes quickly, so the interesting work moves elsewhere. I spend my time on what shapes the product: the art direction, how it is built and on which stack, and the interface, reworked until using it feels obvious.

The code comes fast now. The decisions still take time.

Some projects start with something that keeps getting in my way. Others start by accident: a word, a link, a question that will not leave. Either way, the first thing I write down is who it is for and what they do differently once it exists.

A project gets its name, its logo and its art direction first. And before I fall for a name, I search it: who already ranks for it, whether the domain is free, whether someone typing it would land on me. A name nobody can find is a launch nobody sees.

Then I am picky. Patchwire's logo took seven rounds, and its first four art directions went to the bin.

Check the search results before you fall for a name.

The spec says what ships first, what waits, and what is never stored. The build runs on Claude Code, in small steps with tests at each one. My part is the calls: what to build, what to cut, what is good enough to ship.

Patchwire went from a name to a live product in two days. Preem took four.

A side project dies when it costs too much to keep. So running costs are a design constraint from day one. Patchwire reads its sources every hour inside the free minutes of GitHub Actions, and keeps no accounts and no data about you: nothing to pay for, nothing to leak.

If I can't afford to run it for years, it isn't done.

Shipping teaches what no spec can: what things really cost, which words on the landing page work, how people find it, what they skip. So I put things online early, watch who comes, and fix what matters most.

Right after launch I deleted Patchwire's lockfile reader. Your agent already has your lockfile open: it does not need my site to read it.

Each project leaves a habit behind for the next one. The loop does not change, it just gets faster.