Why Claude Opus 5.5 wants teams to rethink their prompts

Anthropic is advising users of Claude Opus 5.5 to revisit prompt settings and agent workflows, even when existing prompts worked well with earlier versions. The company says the model responds faster than Claude Opus 5, often completes work with fewer tokens, and needs different controls for effort, progress reporting, and safety.

This official document explains that existing Claude Opus 5 prompts should generally remain usable. However, the guide recommends starting at the model’s medium effort setting, rather than automatically carrying over the high setting used with its predecessor.

Effort is now the main control for the amount of internal processing Claude performs. Anthropic says Claude Opus 5.5 always uses thinking, unlike some configurations of Claude Opus 5. Users can select lower effort for quicker and cheaper responses, or higher settings for demanding work such as complex coding, research, and detailed analysis.

The company cautions that higher effort can also mean longer responses and more output tokens. Internal thinking counts toward the maximum token allowance, even when users do not see that thinking. Anthropic therefore recommends setting a sufficiently high token limit for lengthy tasks, particularly autonomous work that involves many tool calls.

More visible progress for long-running work

One of the most practical changes concerns AI agents that work across several steps. Claude Opus 5.5 can send short status notes between tool calls, describing what it found and what it plans to do next. Those updates appear as special thinking blocks, not normal text responses.

That distinction matters for teams building content, research, or automation workflows. A chat interface that displays only regular text may make a long-running task appear inactive, even while the model is working. Anthropic recommends enabling the “updates” display option so applications can show readable progress summaries.

The guide also warns that an agent may end a turn with a progress report while work remains. Systems that interpret any text-only response as completion could stop too early. Anthropic suggests tracking open tasks in a checklist and automatically asking the model to continue when unfinished items remain. It also advises setting limits on those automatic follow-ups to avoid endless loops.

For workflows involving several connected applications, such as email, documents, spreadsheets, and CRM systems, Anthropic recommends instructing the model to search broadly before changing anything. The company says relevant context is often located in files or records that the original request does not mention. This approach can improve task completion, but it requires careful handling of untrusted material found in those sources.

Prompt injection and pasted material

The documentation places particular emphasis on content copied into a user message. A pasted email, web page, or document may contain hidden or misleading instructions. Anthropic recommends wrapping pasted material in application-generated tags and telling the model to treat instructions inside those sections as untrusted unless they match the user’s explicit request.

This is relevant for editorial teams that summarize correspondence, analyze public documents, or process client material. The measure is not a complete defense, Anthropic says, because plain-text tags can be imitated. It should sit alongside other safeguards.

The guide also says Claude Opus 5.5 reads charts, screenshots, and diagrams more accurately than the prior model. For dense visual material, Anthropic still recommends high-resolution images and tools that allow the model to crop, zoom, and inspect source files. For content professionals, that could reduce errors when extracting information from slides, dashboards, calendars, or visual reports.

Finally, Anthropic advises chat applications to remove generic instructions telling Claude to “think carefully” before answering. The company says effort settings are a more effective control and that removing such wording can reduce delay before a response begins.