Process
How an idea becomes a shipped thing. The loop I run, and the side projects that taught me the method.
The loop
Idea
One or two sentences. Enough to argue with.
Spec
Goals, non-goals, and what done means. Written to be grilled.
PRD
The product case and the stories. Open questions listed, not smoothed over.
Technical design
Grounded in the real code, not the remembered code.
Plan
Slices a fresh session can run without the conversation that made them.
Build
One slice, one branch, one draft PR. Delegated to the seat that fits.
Verify
Run the real thing. A green build is not proof.
Ship
Review, then release. The journal entry is the last commit.
Every stage is a document a fresh session can pick up. I sit above the loop: I decide, I review, and I don't type the middle. That's the whole trick of working with agents. The human's job is the parts that need judgment.
Side projects
The same app, three times
A draft pool for Survivor, a bracket for the World Cup, a fantasy league for The Challenge. Each one got the architecture the problem earned: static HTML and a CSV, then a React app with no backend, then real auth, a database, and tested scoring.
modern-record-player
A stack of RFID cards, a Raspberry Pi, and Spotify. Scan a card, the album plays. Registering a new card is a form, not a code change.
card-tracker
A value tracker for 121 women's soccer trading cards, generated by a script and fed by a custom Claude Code skill.
This site
Planned in one document, built by the seats in five slices, each one verified in a real browser before it merged. The repo goes public at launch.