VibeusAnthropic’s Claude Opus 5.5 prompting guide is a useful migration... - Vibeus
Anthropic’s Claude Opus 5.5 prompting guide is a useful migration checklist. It covers effort settings, agent loops, connected apps, visual inputs and frontend defaults. Anthropic says Opus 5.5 generates output tokens more than 30% faster than Opus 5 and tends to finish the same task with fewer tokens.
The guide’s 12 practical takeaways:
1. Give multi-agent work a time budget and report elapsed time against it. That can help teams pace the task, but the budget is advisory, not a hard stop.
2. Progress updates now arrive as thinking blocks. Their text can be empty by default, so an app that renders only text blocks may look frozen.
3. Thinking is always on in Opus 5.5. If an Opus 5 integration disabled it, start at low effort and measure latency and quality.
4. An unattended agent can stop after a text-only progress update. Make the harness check whether the actual task is complete before ending the run.
5. The guide calls out safeguards for biology, cybersecurity and reasoning extraction. Asking the model to print its internal reasoning can trigger a refusal.
6. Mark pasted email or web content with matching tags and a short random ID so the model can distinguish copied text from the user’s own instructions.
7. For work across connected apps, tell the agent to inspect relevant emails, documents, spreadsheet tabs and records before acting, including sources the request did not name. Anthropic reports better completion in its tests, with somewhat more tool calls and tokens.
8. Opus 5.5 reads charts, diagrams and screenshots more accurately, but dense visuals can still benefit from higher-resolution images and cropping tools.
9. In chat prompts, “think carefully” may be unnecessary: Anthropic reports that removing it made replies start sooner without a clear quality drop in its test.
10. For frontend work, naming specific styles to avoid works better than saying only “avoid a generic AI look,” which can just replace one default with another.
11. Effort is the main control. Medium is Opus 5.5’s default and matched or beat Opus 5 at high effort on Anthropic’s coding and knowledge-work evaluations; lowering effort cuts thinking more reliably than prompt instructions.
12. Existing Opus 5 prompts should still work as a starting point; measure before rewriting them.
One cache detail: changing the top-level effort between requests invalidates the prompt cache. The guide says the per-message effort change (beta) keeps it.
