A major behavioral shift in Claude 4.x models: they take your instructions literally and do exactly what you ask, nothing more. Earlier models sometimes added unrequested extras.
If you want thoroughness, ask for it:
Bad: "Review this code"
Good: "Review this code thoroughly. Check for security vulnerabilities,
performance issues, error handling, and edge cases. If you find
problems, suggest specific fixes with code examples."If you want proactive behavior, enable it:
<instructions>
When implementing changes, don't just suggest them — actually make the
changes. Infer my intent and act on it. If something related should also
be updated (tests, docs, types), update those too.
</instructions>If you want creative or “above and beyond” responses:
Don't just answer the question — think about what I'm really trying to
achieve and suggest better approaches if they exist.This also means prompts from older Claude versions may need updating. If a prompt relied on Claude volunteering extra information or going beyond the literal request, you now need to make those expectations explicit.
The upside: This literal behavior makes Claude 4.x much more predictable and controllable. When you give precise instructions, you get precise results.