"Consider several approaches" multiplies your agent's reasoning tokens 2.4–7.4x with no gain in correctness — a preregistered 4,643-run benchmark
Across six large reasoning models, two real agent harnesses and 24 deterministic coding tasks with hidden evaluators, prompt wording alone was shown to multiply reasoning spend without improving success. Asking the model to develop and compare several approaches is the single most wasteful instruction (2.4–7.4x more reasoning tokens across every model); generic "think deeply" cues add 1.6–2.2x; a bounded-efficiency template that states scope, acceptance criteria, and an explicit stop condition is cost-neutral and can halve reasoning. Two builder-relevant kickers: harness choice mattered more than the prompt (identical model-task-prompt triples cost 5–30x more per success under Claude Code than under pi, mostly from larger static prefixes and more turns), and misleading architectural hints in a prompt cost far more than irrelevant prose — so a stale line in your CLAUDE.md is more expensive than a verbose one.
↳ Follow the thread