Getting the most out of Opus 5.5 in Claude and Claude Code
4 hours ago
- Give the entire task in one message with a clear finish line (e.g., 'tests pass') and let Opus 5.5 work autonomously.
- Remove explicit thinking prompts like 'think carefully' since Opus 5.5 always thinks before replying; for simple questions, say 'Answer directly.'
- Add a rule in CLAUDE.md telling Opus 5.5 when to stop and ask for input vs. keep going; use 'continue' if a run stops with 'Want me to continue?'
- For large codebase tasks, ask Opus 5.5 to split work across subagents and check each result.
- For long runs, have Opus 5.5 keep its task list in a file (e.g., TASKS.md) that survives context window compression.
- After a long run, first check what Claude is waiting on from you, then read its summary.
- Use Opus 5.5 to review diffs/PRs; it catches bugs even at low effort with clear explanations.
- For research, ask Opus 5.5 to report what it couldn't find or verify.
- Attach images (charts, diagrams) and ask specific questions; Opus 5.5 reads images and spatial relationships accurately.
- When designing artifacts, list specific design patterns to avoid instead of vague direction.
- In long chats, add an instruction that earlier answers are settled to avoid slowdown from rethinking.
- Be aware of safety flags: if switched to an older model, remove requests to reproduce internal reasoning; use fast mode for back-and-forth work.