Alina Diskova's wroclove.rb 2026 lightning talk on her everyday AI workflow. Every AI request is a loop: prepare input → model returns output → evaluate → iterate. The model is a black box — not a Ruby developer — so the only controllable levers are the input and the evaluation. Recommends providing explicit acceptance criteria and requirements, and splitting work into small, reviewable parts rather than asking for 'the whole ticket'. Worked example: instead of asking an AI to implement a new controller route in one shot, run three loops — (1) as systems analyst, enumerating cases (happy path, not-found, unacceptable); (2) as automation tester, generating an integration test in TDD fashion; (3) as developer, implementing the route against the existing test. The human keeps the code-reviewer role throughout and gets better results with fewer iterations.