Mesurer au lieu de laisser l'IA deviner
When the expected answer is a measurable value (a timing, a silence length, a total, a count of occurrences), you do not ask the agent to guess it: you give it the tool that measures it and require that it calls it. For a video edit, Whisper returns the text and the word timestamps, FFmpeg measures the silences. For books, a connector reads the entries and a script counts. Every correction you make then becomes a rule kept in memory.