Tech • AI • Robotics • Game
Anthropic Revealed Their Secret Guide to Mastering Opus 5.5
8/10TL;DR
Anthropic’s guidance for Opus 5.5 urges users to lower default effort, trim redundant prompting, and give clearer task, context, and completion instructions to improve results while reducing token costs.
KEY POINTS
Default effort has changed
Opus 5.5 now defaults to medium effort, unlike Opus 5, which defaulted to high. Effort determines how much internal reasoning the model uses before replying, and Anthropic recommends starting at medium rather than carrying over older settings. Reusing a higher configuration from earlier models can produce longer turns and more output tokens without clear gains, so the advice is to raise effort only where testing shows a measurable benefit.
Compare effort levels on real tasks
The recommended way to tune effort is to run the same task at different settings and compare the results. Useful checks include whether a response missed an important requirement, whether the recommendation was materially better, and how long the reply took. A longer answer is not treated as proof of better reasoning, making task-specific evaluation more important than defaulting to maximum effort.
Drop vague “think carefully” instructions
Anthropic says broad instructions such as “think carefully before answering” may be unnecessary for Opus 5.5, because the model already adjusts its reasoning and effort is the main control. In testing, removing those phrases made replies start sooner with no obvious drop in quality. More specific prompts, such as asking the model to compare two proposals against requirements and explain trade-offs, are presented as more effective than generic calls for deeper thinking.
Earlier answers may be revisited
Opus 5.5 can reconsider prior answers while processing a new message, even a simple follow-up. That can be useful in research or analysis where new evidence appears, but it may also create unnecessary work. A targeted instruction telling the model to treat earlier answers as settled unless they are explicitly questioned can reduce extra reasoning and save tokens.
Relevant context should be actively gathered
The guidance stresses that the first document in a workflow may not contain the latest or most complete information. In multi-app tasks, deadlines, rules, or updates may sit in emails, spreadsheet tabs, or records not named directly in the request. Asking the model to inspect relevant sources before acting improved completion in Anthropic’s multi-app testing, though it required slightly more tool calls and tokens.
Separate the request from pasted content
When users paste emails, webpages, or other external material into a prompt, that text may include instructions that are not the user’s intent. The recommendation is to clearly mark the boundary: state the job first, then provide the material to read or analyze. That simple separation helps the model distinguish the actual request from third-party content embedded in the prompt.
Silence may be an app problem, not a model problem
Opus 5.5 can generate progress updates that some custom applications fail to display. In those cases, asking the model to “talk more” will not solve the issue if the app is hiding the updates. The practical fix is to ensure the application can receive and show progress messages, and to specify useful update points such as when review begins or when a draft is ready.
A reply does not always mean the task is done
A progress update can end a turn even while work remains unfinished. The guidance recommends defining completion explicitly by listing the expected outputs, such as a report, source list, and summary, and then checking whether each item arrived. If something is missing, users should point to the missing item directly and ask for completion or an explanation of the blocker rather than simply saying “keep going.”
Specificity also matters for design and images
For design work, vague feedback such as “make it less generic” tends to swap one default style for another. Better prompts name concrete elements like background, headline treatment, button shape, and spacing. On visual analysis, Anthropic says Opus 5.5 reads images better than Opus 5, but dense charts and tiny labels still benefit from high-resolution images, close-ups, or crop tools, and the model should be told to say when a detail is unreadable instead of guessing.
CONCLUSION
The guidance around Opus 5.5 centers on a practical shift: rely less on blanket prompting habits and more on precise controls, explicit context, and clear completion criteria. That approach can improve speed, reduce wasted tokens, and produce more dependable results across everyday and automated workflows.
Ask a question