Anthropic Releases Official Prompting Guide for Claude Opus 5.5

Anthropic has released an official prompting guide for Claude Opus 5.5, outlining key changes in agent behavior, safety mechanisms, and performance tuning. The model is reported to be 30%+ faster than Opus 5 and uses fewer tokens, with existing prompts expected to run unchanged.

Agent Behavior and API Changes

• Time Budgets: Adding elapsed time against a budget (e.g., "elapsed 340s / 1200s") helps agent teams pace themselves and finish tasks sooner.
• Thinking Blocks: Progress updates between tool calls now arrive as thinking blocks, which are empty by default. Applications that only display text blocks may appear frozen if not updated to handle these new structures.
• Thinking Control: Thinking cannot be disabled via API plumbing. Users who previously ran Opus 5 with thinking disabled should start at low effort and measure results.
• Turn Termination: Some progress updates now end the turn. Loops that interpret this as task completion may quit prematurely.

Safety and Security Guardrails

• Classifiers: New safety classifiers cover biology, cybersecurity, and reasoning extraction. Users are advised not to ask the model to write out its reasoning in replies.
• Prompt Injection Defense: Wrapping pasted text in tags with a random ID helps the model ignore instructions hidden within copied content.

Performance and Prompting Tips

• Effort Dial: Medium is the default effort level, matching or beating Opus 5 on high effort. Lowering effort cuts thinking more reliably than prompt instructions.
• Cache Invalidation: Changing the top-level effort value between requests invalidates the prompt cache. To vary effort per turn, use the per-message effort change (beta) to preserve the cache.
• Prompt Optimization: Removing "think carefully" from chat prompts made replies start sooner with no clear quality drop. For design tasks, naming exact styles to ban is more effective than generic instructions like "avoid a generic AI look."
• Image Analysis: Opus 5.5 reads charts, diagrams, and screenshots more accurately, though higher-resolution images and crop tools still help with dense visuals.
• Context Gathering: Asking the model to check emails, docs, and records across connected apps before acting improved results.