Most of what I do with AI is not prompting. It is the layer around the models: what runs, against which repo, under what written goal, and how anyone knows afterwards whether the work happened.
That layer runs many unattended coding sessions in parallel across about seventy tracked pieces of work, with state in files rather than in a session’s memory, and every report compared against what the session was asked to do before it is believed. Twenty-six scheduled jobs run alongside it, all routed through one gateway, so work goes by task to a local model or a cloud one.
What comes out is ordinary working software: two plugins published on npm, a public statistics site covering more than 200,000 harbour visits that regenerates every night, telemetry running on hardware at sea.
The part worth paying for is the second half. On 2 August a job wedged for four hours and eleven minutes while every thirty-minute fire reported ok — correctly, because the wrapper exited zero. The night before, the same defect cost twenty hours and forty minutes. Both times the scheduler said fine and nothing had moved.
So progress is measured at the artifact the job should change — a row count, a file date, a published timestamp — never at its own exit status. Same rule elsewhere: a wrong number is worse than no number, and a deploy is verified by looking at the live system, not by reading the report on it.
The model is confident either way. The engineering is knowing what to check.