/ Article
Vague prompts produce vague products. Notes from steering experiments on how LLM agents actually behave when you give them room to improvise.
I spent a week whispering at agents — soft suggestions, half-constraints, polite preferences. The outputs looked fine in isolation and fell apart as a product.
Ambiguity gets amplified
Agents amplify ambiguity. If you don't say what "done" means, they invent a version of done that is locally coherent and strategically wrong.
Louder steering
The fix wasn't better vibes. It was louder steering: explicit acceptance criteria, forbidden patterns, and review rituals that catch drift before merge.