Anthropic: delete the “double-check your answer” lines from your Opus 5 prompts

Anthropic’s new prompting guide for Claude Opus 5 says prompts written for Opus 4.8 carry over without edits. But instructions like “add a final verification step” or “re-check your answer” now trigger extra cycles for nothing. Opus 5 verifies its own work, and stripping those lines cuts token spend with no quality hit.

Opus 5 also answers at greater length than previous Opus models by default. The effort parameter controls how much the model thinks, not how long the visible response is, so brevity has to be requested explicitly in the prompt.

Reasoning is on by default and can only be turned off at effort high or lower. With it off, the model sometimes prints a tool call as plain text instead of executing it. Anthropic recommends pulling “do not reason” rules out of the system prompt entirely: they make internal XML tags leak more often.