How we work

Software built with agents, in the right order. Five steps, always the same, and three things that stay with people.

1. 01Conversation and a short spec

   We start by asking until no doubt remains: what counts as done, what is out, what must not change. The result is a one- or two-page specification with the decisions closed. A spec with open questions does not go to implementation.

2. 02An executable plan

   The spec becomes a plan of small tasks. Each task names the files it touches, the test that proves it and how it is verified. Someone reading the plan without context can execute it.

3. 03Agents execute, with a record

   Agents follow the plan task by task: write the test, watch it fail, write the minimum that makes it pass, commit. Deviations from the plan are recorded with the reason, for whoever comes next.

4. 04Review with the spec alongside

   A person reviews intent, not syntax: is what is there what was asked for, no more and no less? A reviewer with no prior context reads the whole at the end, because the author has blind spots.

5. 05Verification before claiming done

   "It should work" is not a state. Before saying it is done we run the tests, the lint, the build and, when there is an interface, screenshots and Lighthouse. Only what ran and passed counts.