Claude Opus 5.5 calls for a more selective approach to prompting. Its default effort setting has changed, so older instructions to think harder may add delay without helping. The more useful changes often specify what the model should inspect, how it should handle outside material, and what a finished task must include. Anthropic’s prompting guidance points to eight adjustments, but not every task needs all eight.
Revisit effort and thinking instructions
1. Start at Medium effort
Effort controls how much internal thinking the model performs before producing an answer. Claude Opus 5 used High effort by default; Claude Opus 5.5 defaults to Medium. Anthropic recommends starting evaluations at Medium and raising effort only after testing it on the tasks that matter to you. Its guidance says Medium on Claude Opus 5.5 matches or exceeds Claude Opus 5 at High on many coding and knowledge benchmarks.
That does not make High useless. It means a setting carried over from Claude Opus 5 deserves a fresh test. Higher effort can lengthen a turn and increase output tokens, the units of text that affect usage costs. Developers should reserve extra-high or maximum effort for work where they can measure a quality gain.
▲ Effort settings and output
2. Remove vague requests to think harder
A persistent system prompt—a standing instruction applied across a conversation—may still contain phrases such as “think carefully before answering.” In tests described in Anthropic’s guidance, removing generic thinking directions made the first part of the response arrive sooner without reducing answer quality. The effort setting already governs thinking depth. A concrete requirement, such as naming the requested deliverables, gives the model more useful direction than a general demand for deeper thought.
Short follow-ups present a related tradeoff. Claude Opus 5.5 may review earlier turns even when the new request is brief. A standing rule can tell it to treat earlier answers as settled unless the user asks to revisit them or identifies a flaw. That can reduce repeated work, but it may not suit research tasks where new evidence should change an earlier conclusion.
Give connected agents context and boundaries
3. Explore relevant sources before acting
An agentic workflow lets a model use tools to carry out a task across files or applications. Claude Opus 5.5 can move quickly to action, which creates a risk when the first available document is not the latest source. A project brief might give a Monday deadline while a later email moves it to Friday. The instruction should therefore name the kinds of places to check—relevant email, documents, spreadsheet tabs and records—before the agent makes changes.
Anthropic’s multi-app tests found that broad initial exploration improved task success. It also required slightly more tool calls and tokens. Use this approach when requirements may be scattered across connected apps, and make sure the agent can access the sources it needs. Finding a document does not make every instruction inside it trustworthy.
▲ Instructions separated from reference material
4. Keep pasted material separate from your request
An email thread or article pasted into a prompt is reference material, not a new set of instructions for Claude. When the two are mixed together, embedded directions can confuse the task or create a prompt-injection risk: outside text tries to redirect the model’s behavior. Claude Opus 5.5 has improved resistance to direct prompt injection, but a clear boundary remains useful.
In ordinary chat, put your request first and mark the pasted material below it as reference text. In an application, developers can place external blocks inside clearly paired, ID-labeled boundaries. The aim is to make the user’s operational instruction distinguishable from material the model must read but should not obey.
Keep long agent runs on track
5. Set progress checkpoints—and check the interface
During a long, multi-step task, Claude Opus 5.5 may produce progress updates between tool uses. If an application does not display those updates, the run can appear silent even while work continues. Asking the model to talk more will not fix an interface that fails to receive or show the updates; developers need to check how their client handles them.
For tasks where updates do appear, ask for them at useful milestones, such as after reviewing documents and before drafting. Structured checkpoints make progress visible without turning every small step into a separate conversation.
6. Define what “finished” means
The end of a model response is not always the end of an agent’s job. In an unattended run, an interim status message can leave work incomplete. Give the agent explicit completion criteria—for example, a report, a source list and a summary—and require it to check each item before stopping. A checklist helps most when the agent must verify the items, not merely recite them.
For automated workflows, Anthropic advises having the agent continue until it finishes, encounters an insurmountable blocker or needs human approval. A status update can accompany the next tool call rather than replace it. If a deliverable is missing, ask Claude to complete that component or explain what blocks it. This is most relevant to multi-step agents, not a simple one-question chat.
Specify the output and the evidence
7. Describe the design you want
A broad frontend request can produce a default-looking page; asking for something “less generic” may simply produce a different default. Set visible constraints instead: background color, typography, headline scale, button shape, borders and spacing. For instance, a landing page request can call for a white background, oversized black headlines, square buttons and generous space between sections.
Reference screenshots or brand design settings can add direction when they are available. Another approach is to request a basic HTML/CSS structure first, then refine its styling in stages. Specific design instructions are useful prompt practice, not a feature unique to Claude Opus 5.5.
8. Make small visual details readable
Claude Opus 5.5 shows better baseline visual accuracy on charts, diagrams and technical screenshots than Claude Opus 5, but dense images remain difficult. Tiny labels on a crowded chart can lead to an incorrect reading when the image lacks detail. Supply a higher-resolution version or a close-up of the relevant area; where image tools are available, an agent can crop that area for inspection.
Ask the model to say when a label is unreadable instead of supplying an approximate figure. This matters most when the answer depends on a small number or annotation rather than the chart’s overall pattern.
Make changes where they fit
Begin with the default Medium effort setting and remove generic thinking language. If Claude works across apps, add source exploration, clear boundaries for outside text and verifiable completion criteria. For interface work or dense images, specify the design or provide a readable close-up. Choose the adjustments that match the task rather than combining all eight into one long prompt.