Anthropic releases Opus 5.5 usage guide

Anthropic recommends assigning tasks with defined completion criteria and stop points. Users should avoid manual "think carefully" prompts, as the model now manages its own reasoning process, and utilize sub-agents for large-scale work.

Key recommendations include providing full task context in one message, removing "think carefully" phrases, and using CLAUDE.md to define interaction rules. For complex tasks, users should employ sub-agents and verify reports before final review.

The guide suggests providing clear, verifiable success states, such as passing tests or migrating endpoints. Opus 5.5 excels at maintaining long-running tasks without constant supervision when the end goal is explicitly defined.

Users should remove "think step by step" prompts, as Opus 5.5 automatically adjusts its reasoning depth. For Claude Code, users can adjust effort settings, and they can provide additional instructions during execution without restarting the task.

To avoid generic design outputs, Anthropic advises listing specific patterns or styles to exclude. By iteratively adding forbidden elements to a list, users can narrow the model's output toward their desired aesthetic.

Users should define clear stop rules in CLAUDE.md to manage autonomous behavior. The model should be instructed to proceed without input for routine tasks but pause for high-risk actions like data deletion or repository changes.

For large-scale audits or migrations, the guide recommends delegating tasks to sub-agents. The main session should act as a coordinator, verifying the evidence provided by each sub-agent before finalizing the results.

Maintaining a TASKS.md file allows users to track progress through context summaries. Upon completion, the model should be instructed to provide a structured report covering pending decisions, changes, and findings.

Anthropic suggests having Opus 5.5 perform initial reviews of its own work. Users can prompt the model to identify potential bugs, explain reasoning, and explicitly mark areas where information remains unverified.

Opus 5.5 shows improved accuracy in interpreting visual data like charts and diagrams. Users can request full documents or spreadsheets rather than summaries and should use project instructions to maintain focus in long conversations.

Opus 5.5 includes updated safety guardrails. If a message triggers a transition to an older model, users can manually re-select Opus 5.5 or adjust their prompt. Direct requests for internal reasoning are blocked, so users should ask for specific explanations instead.

A new "fast mode" is available for Claude Code, intended for interactive, short-turn tasks. While it provides quicker responses, it incurs higher costs and is less efficient for long-running autonomous tasks compared to standard mode.

The user's role is evolving from providing step-by-step instructions to designing task boundaries and verification protocols. Success depends on clear stop conditions and effective use of CLAUDE.md to manage the model's autonomy.

The full guide and additional details regarding Opus 5.5 optimization are available on the official Claude developer blog.