I direct, the agents type.
I build my projects the way a small studio works. I write the plan and the briefs, coding agents build in parallel, each on its own branch, and I review, play and test everything before it merges. The judgment stays mine; the typing is shared.
- 1,982commits since the first one in July 2026
- 467pull requests merged
- 265test files guarding the game
- 42hand-built walks in and around Barcelona
Path of Leash Resistance, as of November 2026.
1. Write it down first
Every project has a PROJECT.md with its design pillars and roadmap, and anAGENTS.md that any coding agent reads before it touches anything: the tech stack, the conventions, the game's voice, and the rules that are not up for negotiation. When an agent gets something wrong, the fix goes into those files as well as the code, so the next agent doesn't repeat it.
2. One brief per agent, one worktree per brief
Work goes out in batches. Each piece gets a written brief: a shared preamble with the house rules, a shared brief for that kind of work (say, building a new walk), then the specifics. Each agent works in its own git worktree on its own branch, so a batch of four new walks, or a polish pass across every district of the city, can be built at the same time without anyone stepping on anyone else.
The briefs carry lessons forward. A batch prompt starts with what earlier batches got wrong: a test that assumed its walk was the last one, poles that all grew street lamps. Those mistakes don't come back.
3. Nothing merges unchecked
Every pull request has to show its work before it can land:
- The walk's own tests and the full suite pass, plus a headless smoke run of the game.
- A scripted playtest walks the level from start to finish and logs anyone stuck or inside a wall.
- Screenshots of every part of the route, in every weather and at night.
- A feel card: three to five things to try by hand and what should happen. I play them.
- A changelog entry, checked by a script.
4. Release like a team would
A version tag builds the web and Windows versions in CI, pushes them to itch.io, and packages the Microsoft Store build. stemma ships the same way, with a pinned release lock so CI tests exactly what users get.
Why work this way
The hard part of software was never typing. It's deciding what to build, noticing what's wrong, and keeping a large codebase coherent as it grows. Agents make the typing cheap, which makes those skills matter more, not less. This is how one person ships a game with 42 levels, eight chases and four bosses in a few months, alongside a day job.