Anthropic is encouraging users to change how they interact with its Claude Opus 5.5 model, arguing that prompting techniques developed for older generations of AI systems can now make the experience slower and less efficient. The company says users should stop relying on instructions such as “think hard” and “think carefully” because the latest model already adjusts its reasoning automatically based on the complexity of a task and the selected effort setting.
The recommendations are part of Anthropic’s updated guidance for getting more out of Claude Opus 5.5. The company also wants users to provide complete instructions upfront, request finished documents instead of preliminary drafts, share charts and diagrams directly, and connect Claude to their work files and applications. The broader message is that users should move away from managing every individual step and instead give more responsibility to the AI model.
Why Anthropic Wants Users to Change Their Prompting Habits
Prompt engineering has changed significantly as AI models have become more capable.
Earlier reasoning models sometimes benefited from explicit instructions telling them to spend more time thinking through a problem. Phrases such as “think step by step,” “think carefully” and “think hard” became common ways of encouraging models to produce more considered answers.
Anthropic says that approach is no longer necessary for Claude Opus 5.5.
According to the company’s testing, the model already reasons automatically and determines how much effort a task requires. Its reasoning depth is influenced by the selected effort level and the complexity of the request.
Anthropic found that removing phrases such as “think hard” and “think carefully” helped responses begin faster without a measurable reduction in answer quality.
This means users may be adding unnecessary instructions that do not improve the final result.
The company is not suggesting that users should stop giving detailed instructions. Instead, it wants them to focus on explaining the task, desired outcome and constraints rather than telling the model how to think.
Start With Medium Effort Instead of High
Another major recommendation concerns Claude Opus 5.5’s effort settings.
Anthropic recommends starting with Medium effort, which is the model’s default setting, rather than immediately selecting High or the more expensive higher-effort configurations.
The company says its internal evaluations found that Opus 5.5 at Medium matched or outperformed its predecessor, Opus 5, running at High effort on coding and knowledge-work tasks.
Higher effort settings can increase reasoning time, token consumption and operating costs. However, they do not guarantee better results for every request.
Anthropic recommends moving to High when Medium effort fails to deliver the required quality.
The distinction matters because effort settings control how much computational work the model performs. Users who automatically select the highest available setting may be spending more without receiving meaningful improvements.
For developers using the Claude API, Anthropic also recommends testing different effort levels against their own workloads instead of assuming that settings used for an older model will remain optimal.
Give Claude the Complete Task in One Message
Anthropic also wants users to change how they structure their requests.
Instead of providing instructions one at a time, the company recommends describing the complete objective in the initial prompt.
A useful request should explain three things:
- The goal: What needs to be accomplished?
- The expected output: Should Claude produce a report, spreadsheet, presentation, code change or another finished deliverable?
- The stopping point: When should Claude ask questions, request approval or stop working?
According to Anthropic, Opus 5.5 is better at handling multi-step assignments than its predecessor. It can manage more of the intermediate planning and follow-up that users previously had to coordinate themselves.
For example, rather than asking Claude to research a subject, waiting for the answer and then separately requesting a report, users can explain the research objective and request the finished report in one instruction.
This approach can reduce unnecessary back-and-forth and help the model understand the intended result from the beginning.
Ask for Finished Documents Instead of Drafts
Anthropic says users should also reconsider how they request professional documents and other deliverables.
When someone needs a spreadsheet, presentation or report, the company recommends asking for the completed file directly instead of requesting an outline or preliminary draft first.
Opus 5.5 can increasingly move from an initial request to a finished deliverable in a single pass, according to Anthropic.
The recommendation reflects a broader shift in how AI assistants are being used. Rather than simply generating text in response to individual questions, advanced models are increasingly expected to complete entire assignments.
This can be particularly useful for business users who regularly prepare financial spreadsheets, research reports, presentations and project documents.
However, users should still review important deliverables for factual accuracy, calculation errors and formatting problems before relying on them.
Share Charts and Diagrams Directly
Anthropic is also recommending that users stop manually converting visual information into text whenever possible.
Older AI systems sometimes struggled to interpret complex charts, diagrams and tables. Users consequently developed workarounds, including manually typing chart values into prompts.
The company says Opus 5.5 has made substantial improvements in visual understanding.
Its guidance recommends attaching the original chart, screenshot, diagram or table rather than transcribing the information manually.
Anthropic reports that even at the lowest effort setting, Opus 5.5 read values from dense charts more accurately than Opus 5 did at its highest effort setting in the company’s evaluations.
This improvement could make the model more useful for financial analysis, business research, technical documentation and other tasks involving visual data.
Users can provide the original material and ask Claude to extract information, identify patterns or explain what the data indicates.
Be Specific About Design Preferences
Anthropic has another recommendation for people using Claude to create websites, interfaces and other visual designs: describe the exact design patterns they want to avoid.
A general instruction such as “avoid a generic design” may not produce the intended result. The model might simply replace one common design style with another.
Instead, Anthropic recommends naming specific elements that should not appear.
For example, users can explicitly reject cream-coloured backgrounds, pill-shaped buttons, excessive gradients or other visual choices that do not match their preferences.
The company also encourages users to provide their own design guidelines, examples and existing work.
Sharing a brand guide or pointing Claude towards previously completed presentations can help the model understand the visual and editorial standards it needs to follow.
Connect Claude to Work Files and Applications
Anthropic considers connecting Claude to external work systems one of the most useful ways to improve its results.
When users connect relevant files and applications through supported connectors, Claude can retrieve information that is not included in the original prompt.
For example, someone preparing a business report may have relevant information spread across documents, spreadsheets and other work applications.
Without access to those sources, the model can only work with the information supplied directly by the user. With appropriate connectors and permissions, it can gather additional context before producing the requested output.
Anthropic says this can lead to more tailored results because Claude can access information that would otherwise need to be manually collected and pasted into the conversation.
Connector availability and permissions depend on the user’s plan and organisational settings.
The Bigger Picture
Anthropic’s recommendations reflect a broader transition in AI prompting. As models become better at reasoning, interpreting images and completing multi-step assignments, users may no longer need to manage every stage of the process through separate instructions.
The company is effectively encouraging users to move from conversation-based prompting towards outcome-based delegation. Instead of explaining every action, users should describe what they want completed, provide the necessary context and allow the model to determine the steps.
This also has implications for the cost of AI adoption. Selecting unnecessarily high effort settings or repeatedly prompting a model to think harder can increase resource consumption without improving results. Better task descriptions and appropriate effort settings could make advanced AI systems more efficient.
Looking Ahead
Anthropic’s guidance suggests that the best way to use newer AI models may be different from the techniques that helped users get better results from earlier systems. As reasoning becomes an integrated capability, prompt engineering is increasingly about defining objectives, supplying relevant context and specifying the desired output rather than repeatedly instructing a model to think more carefully.
For Claude users, the practical recommendation is straightforward: start with Medium effort, remove unnecessary thinking instructions, provide complete tasks upfront and use available files and visual inputs. As Anthropic and its competitors continue improving their models, prompting practices are likely to evolve alongside them.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.



