Anthropic has published guidance on optimizing use of Opus 5.5, its latest Claude model, highlighting techniques for handling extended tasks in Claude apps and Claude Code.
According to the guide, Opus 5.5 performs best on long, multi-part work—its largest improvements over prior Opus models come on multi-step tasks like migrating code changes through large repositories until tests pass. Early testers ran long coding tasks for hours with minimal oversight.
Key recommendations include:
Provide clear finish lines. Give the entire task in one message and define what “done” means—“every endpoint uses the new client, the old client is deleted, and the test suite passes,” for example. This helps the model know when to stop.
Remove explicit thinking prompts. Delete lines like “think carefully” or “think step by step” from prompts and saved instructions. Opus 5.5 thinks before every reply and decides how much automatically. Removing these lines made responses start sooner with no clear quality drop in testing.
Interrupt mid-run with updates. Runs are longer now, so restarting is costlier. Users can type follow-up messages while the model works, such as “Also keep the old endpoint names as aliases.”
Use design constraints for UI work. When requesting pages or apps, list specific design patterns to avoid rather than vague directions. For example: “Don’t use a cream or off-white background, italic accent words in headings, numbered section labels, or pill-shaped buttons.”
Set stopping rules in CLAUDE.md. Add instructions about when the model should ask for input versus continue working independently. The guide suggests: “When a step doesn’t need my input, keep going. Stop and ask only when you can’t continue without me, or before anything destructive.”
Delegate parallel work. For audits or migrations across large codebases, ask Opus 5.5 to split work across subagents and check results.
Maintain task lists in files. For long runs, request the model keep a checklist in a file like TASKS.md. Long runs fill the context window, causing Claude Code to summarize older messages, but files persist.
Review diffs before humans do. One tester reported Opus 5.5 at lowest effort caught more bugs than Opus 5 at high effort with fewer false alarms.
The guide also notes improvements in reading charts and diagrams, catching detail inconsistencies in long documents, and generating spreadsheets requiring less post-generation editing.
Key facts
- Opus 5.5 excels at multi-step work and can run long coding tasks for hours with minimal oversight
- The model thinks before every reply and does not need explicit prompting to think carefully
- Users can interrupt runs with follow-up messages while the model is working
- Setting clear finish lines and stopping rules in CLAUDE.md helps steer long tasks
- Opus 5.5 reads charts, diagrams, and screenshots more accurately than Opus 5
- The model can coordinate parallel subagents for large audits and migrations