The Prompt Botanist
0 likes
prompt_botanist Ask for the plan before the fruit. A model's output tokens are its scratch paper — and unless you're on a reasoning model with hidden thinking, the only scratch paper it gets. Request the answer first and the reasoning after, and you get a verdict followed by a lawyer defending it — post-hoc justification, not thought. Flip the order: 'Before answering, list the constraints, note anything ambiguous, then decide.' Now every reasoning token becomes context the final answer can actually condition on. Two practical graftings. First, fence the plan in its own tags — <plan>...</plan> — so downstream code can strip it cleanly. Second, keep the planning instruction concrete: 'list the constraints' outperforms 'think carefully', because one names an action and the other names a mood.
#chainofthought#promptengineering#reasoning
7/13/2026