Anthropic tells Opus 5.5 users to stop asking it to think
A command that strips old prompting habits out of a project cut costs a further 9 per cent in the company's own test, and is spreading in three languages.
Lance Martin said on Wednesday he had updated the Claude Code command for prompt auditing with guidance for Claude Opus 5.5. It reads a project's skills, CLAUDE.md, agent files and prompts, then proposes removing instructions that Anthropic says hold newer models back. His post had 2,642 likes and 171,263 views.
Anthropic's own post, written by Martin, lists what the audit looks for. Verification rituals such as double-check your work. Thoroughness boosters such as be maximally thorough. Mandatory step-by-step procedures, scratchpad scaffolds, few-shot examples tuned to older failures, and settings left over from earlier Claude generations.
The company put a number on it. On a customer support benchmark, moving from Opus 4.8 to Opus 5.5 cut cost by 18 per cent on its own, and running the audit cut a further 9 per cent while adding about 2 percentage points of accuracy. The measurement is Anthropic's, on a benchmark it chose.
Addy Osmani made the same argument in a guide published on Tuesday. Opus 5.5 always thinks before it replies and decides how much, he wrote, so lines such as think carefully should come out. He said removing one such line made replies start sooner, with no clear drop in quality.
The advice is travelling. A post relaying the command drew 2,452 likes, a Japanese post 1,492 and a Portuguese one 1,889, all within a day. Whether the audit helps a given project may depend on how much old scaffolding that project is carrying.