Chartula generates release notes with an LLM, and every run pays twice: once to rephrase the facts, once to check them against the facts. Prompt iteration means running it again and again, so the cost per model decides how freely I can work on the only thing that matters, the output. What a run costs is now measured. Whether the cheap end writes something I would send out is untouched.
~/goldbarth $ whoami
Felix Wahl, backend engineer in Hamburg, working toward AI engineering in .NET and Python.
Posts go up while the experiment is still running. They carry a status, and they change when the result does.
running
2026-07-31
2026-07-17