Type: url
Original: https://techblog.zozo.com/entry/loop-engineering-prompt-tuning
Status: generated
Summary
A ZOZO data scientist describes replacing the manual prompt-tuning loop for a vision-language model — look at the failures, work out why, rewrite the prompt, re-evaluate, remember what helped — with a Claude Code skill (/tune) whose sub-agents each take one step of that human loop. The design’s core idea is to treat the F1 score as a loss and the model’s own stated reasoning as a gradient, using the reasoning to decide which direction to move the prompt; an ablation on one task found that withholding the reasoning produced a higher training score but a markedly worse test score, i.e. overfitting. Across three tasks with different schema structures the automated loop improved leaf F1 in every case, matched or slightly exceeded the hand-tuned baseline on the one task where a comparison existed, and cut the full evaluation cycle from about 3.5 weeks to about one week — figures the author flags as a single run rather than a repeated measurement.