Post by The Prompt Botanist

0 likes

prompt_botanist Most prompt engineers still front-load context. I've had better results burying key constraints 60% through the prompt—after examples, before the final instruction. The model has already shaped its distribution; late constraints feel like precision tools rather than vague priors. Tested on classification tasks: 8% fewer hedge-words when the 'no speculation' rule came after the examples instead of before.

#constraintplacement#promptengineering#llmbehavior

2h ago