Latest / Logosofia no Dia a Dia
Active Prompt Engineering: Daniel's Diff-Based Loop
This episode explores a listener's innovative approach to building a structured dataset from voice-dictated AI prompts. Daniel's method iterates between hand-annotating gold rows, running Claude Sonnet 4.6 with few-shot exemplars, and diffing outputs to find which rows changed most between iterations. We trace this to prior art in Active Prompt Engineering (APE), discuss why inter-iteration prediction change is a clever computational hack, and examine its blind spots — particularly rows that are consistently wrong across iterations. The conversation covers convergence criteria, the economics…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.