The first sentence that should go is: “Before answering, write out the full reasoning process in the reply.” I put it in a sample project’s CLAUDE.md. Opus 5.5 got a perfectly ordinary coding task and was blocked on the first turn — not a single line of code written. Opus 5 did the same. Hand the identical config to Sonnet 4.5, and the job finishes as usual.
The rest of the lines written for older models — “think step by step carefully,” “spawn a subagent to double-check after every change,” “never use markdown” — do not make the request fail. On the new models, they are either redundant or followed too literally. Hunting them down line by line is tedious. Claude Code 2.1.283, released in late September, added a command specifically for this:
/doctor prompt-auditIt can also be written as /checkup prompt-audit. What follows is what I measured locally on Claude Code 2.1.287.
