rogetlabs

How I work

AI writes most of the code. I decide what gets built and what stays.

The loop: spec, build with agents, review and test, ship, then lessons feed back into the spec.Specwhat and how to checkBuildagents, small stepsReviewtests and diffShipone change at a timelessons file: every mistake becomes a rule
The loop I run on every project. Tools: Claude Code with custom sub-agents, Make, plain files.

  1. Spec before code

    I write what it should do and how I will know it works. The agent builds from that, not from a chat.

  2. Small steps I can check

    One change, one test, one commit. If I cannot review it in a few minutes, the step is too big.

  3. Keep a change only if a test says so

    In the forecasting model, calibration stayed and an xG blend was dropped. The backtest decided, not my preference.

  4. A lessons file the agent reads

    Every mistake becomes a rule. Each session starts by reading them, so the same bug does not come back.

  5. Say what AI wrote

    AI writes most of the code. I own the architecture, the trade offs and the tests, and I can explain all three.

See the projects