Claude Clue #21: Say What You Mean. No, Really.

Claude Opus 4.7 is more literal than 4.6, and Anthropic flagged it directly in the release announcement: "where previous models interpreted instructions loosely or skipped parts entirely, Opus 4.7 takes the instructions literally." The upside is precision—fewer surprises, less drift, better behavior in pipelines and structured extraction. The catch is that prompts and skills you tuned for 4.6 may suddenly land differently. If your prompt was leaning on the model to fill in the blanks, 4.7 will probably stop filling them in. Anthropic's migration guide is worth a slow read; the literalism note lives under Behavior changes, not Breaking changes.

A code-review prompt that used to read "review this for issues" might come back lighter than you expect. 4.7 reviews for issues, period, and won't extrapolate to security, performance, and style unless you ask. Tighten it to "review this code for security vulnerabilities, performance problems, and style inconsistencies, with specific line references" and you'll get the old behavior back, plus better. Same trick on skills and system prompts: anywhere you implied something, name it. There's a bonus tip in Anthropic's own guide that ties back to Episode 58 and the Boss Rush experiment—positive examples of what you want beat negative instructions about what to avoid. "Respond like this" lands harder than "don't do that."

Roscoe found it, Claude drafted it, I told it to stop assuming what I meant. It complied, literally.

HT: Roscoe Rawson