Anthropic says Claude Opus 5 works best when you stop micromanaging the prompt

CodeByPoonam · x · 2026-07-28

Anthropic published the official Claude Opus 5 prompting guide.
The guide says Opus 5 is tuned for complex agentic coding and large, end-to-end tasks, especially multi-file refactors and bug-finding.
Several prompting habits from earlier models are now counterproductive:
don’t ask it to “double-check” or “re-check” by default, because it already self-corrects;
in code review, letting it flag everything can work better than trying to narrow its focus too early;
to control length, ask for shorter outputs directly, rather than lowering effort;
subagents are useful for big jobs but can be overkill for small ones.
The broader takeaway: some old prompt hand-holding now gets in the way because the model is more capable out of the box.

More from coding & agent