Getting the most out of Opus 5.5 in Claude and Claude Code

Opus 5.5 works well with Claude apps and Claude Code, offering better performance on longer tasks and providing clearer results.

When starting a session with Opus 5.5, give the entire task in one message, including a clear finish line like "the tests pass" or "every endpoint is migrated." This allows Opus 5.5 to work on multi-step work more efficiently than its predecessors.

Avoid including "think carefully" or "think step by step" in your prompts, as Opus 5.5 already thinks before responding and decides the level of thought needed. Simply state "Answer directly" for straightforward questions or adjust the "effort" setting in Claude Code for more or less thinking.

If you need to provide additional information during a run, you can type a follow-up message while Opus 5.5 is working. When requesting a page, app, or artifact, list any design habits you want the model to avoid. A specific list of patterns is more effective than a generic instruction.

Opus 5.5 will inform you of stops in the process, such as "Want me to continue?" or offering choices. You can respond accordingly, and the model will follow your instructions. For destructive commands or risky actions, keep your permission prompts enabled.

For large codebases or audits, ask Opus 5.5 to split the work among subagents and track the progress in a file. This allows you to monitor the status more easily than scrolling through the chat history. When the run ends, review the file for any pending decisions or changes that require your approval.

Look for specific feedback from Opus 5.5, such as "Mark anything you couldn’t confirm, and say where you looked." This helps identify areas where the model couldn’t find or verify information, making it easier to address them.

For image-based tasks, attach the relevant charts, diagrams, screenshots, or slides to your query. Opus 5.5 reads visual content more accurately than its predecessor and can provide specific insights based on the provided images.

Give Opus 5.5 a detailed plan or report and ask it to identify any mistakes. The model is better at finding errors in visual content and can also point out areas where it couldn't confirm or check information. This can be particularly useful for research and analysis tasks or for reviewing pull requests before a human reviewer.