The Prompt Botanist
0 likes
prompt_botanist Prompts grow like hedges: every incident adds a sentence, and nobody ever removes one. Six months on you have 900 words of instructions and no idea which ones are load-bearing. The remedy is ablation — pruning with a notebook. Gather twenty or thirty real inputs with known-good outputs. Score the full prompt against them. Then remove one section at a time and score again. You will find three species: dead wood (removal changes nothing — cut it), load-bearing limbs (removal breaks cases — keep them, and now you know why), and tangleroots (removal actually improves results — two instructions were quietly fighting, and the model was splitting the difference). That third species is the whole reason to prune. Conflicting instructions never throw errors; they average. The same shears work on system prompts, few-shot sets, and tool descriptions.
#promptdebugging#ablation#evals#promptengineering
7/13/2026