Anthropic’s prompting guide for Claude Opus 5.5 advises developers to test multiple effort levels instead of reusing the setting used in Claude Opus 5. It also suggests chat applications on Opus 5.5 consider removing system-prompt lines instructing Claude to think carefully before responding.
Opus 5.5 launched Sept. 22. The guide says Opus 5 prompts should still work well without edits, and that the company’s Opus 5 guidance is a reasonable place to start.
When the company tested this in a chat product, removing a think-carefully line made replies start sooner, with “no clear decline in the quality of the reply.”
Opus 5.5 Defaults To Medium Effort
Opus 5.5 runs at medium effort by default, one level below the high default on Opus 5.
Thinking can’t be turned off on Opus 5.5, and requests that try get an error. Opus 5 allowed that at high effort or below. The guide calls effort the first setting to adjust when balancing quality against speed and cost.
In Anthropic’s coding and knowledge-work evaluations, Opus 5.5 at medium effort matches or exceeds Opus 5 at high effort.
To cut thinking, the guide says to try lower effort settings before prompt changes. Reserve xhigh and max for work where you’ve measured a quality gain.
Anthropic Suggests Reconsidering “Think Carefully” In Chat
The guide’s reason for suggesting chat apps reconsider think-carefully lines is that Opus 5.5 determines how much to think. Effort is the main control.
Anthropic’s own docs include that kind of line. For Claude Opus 4.7, when effort has to stay low to keep responses quick, its page on effort suggests adding guidance such as “This task involves multistep reasoning. Think carefully before responding.”
Agent Guidance Covers Time Budgets And Pasted Text
For agent teams, the guide recommends a time budget if you can estimate how long the task should take. The guide says to have your harness add elapsed seconds against that budget at the end of each message it sends back to the model.
Anthropic’s tests have shown that small agent groups, when using time signals, finish research tasks quicker than a single agent working without signals. Even under time budgets, these groups managed to achieve answer quality comparable to solo agents. The guide says the budget is advisory and doesn’t stop the model at the limit, so keep your own timeout if you need a hard stop. It also says to check answer quality on your own tasks, because the model might search and verify a little less under pressure.
For users who input text from sources like emails, the guide suggests using tags with a random ID and adding a system note on how to manage tagged text. This approach encourages careful handling and, since tags are plain text that can be copied, they provide only one layer of protection against prompt injection.
5 Claude Opus 5.5 Prompting Tips Selected By SEJ
SEJ’s selection of what Anthropic says to retest, consider removing, and add when prompting Opus 5.5
-
Retest effort settings
Don’t assume your Claude Opus 5 setting is still the right one for Opus 5.5.
-
Lower effort before prompting for less thinking
Anthropic says effort is the first setting to adjust when balancing quality, speed, and cost.
-
Consider removing ‘think carefully’ in chat prompts
In Anthropic’s chat-product testing, removing that instruction made replies start sooner.
Anthropic said there was ‘no clear decline’ in reply quality.
-
Use time budgets for predictable multiagent tasks
For multiagent workflows with a sensible task estimate, have your harness report elapsed time against the budget.
The budget is advisory. Keep your own timeout if you need a hard stop.
-
Tag pasted text from outside sources
Anthropic recommends marking pasted content and telling Claude how to treat it.
It describes this as one guardrail against prompt injection, not a full defense.
Anthropic says existing Opus 5 prompts should still work well. These are Opus 5.5-specific tuning suggestions.
Why This Matters
Several sections of the Opus 5.5 guide ask developers to re-check setup carried over from older models, such as workarounds for reading charts and screenshots, or instructions written for running with thinking off. The Fable 5.1 guide, which Roger Montti covered earlier this month, asked developers to revisit formatting rules.
An app that never sets effort runs at medium on Opus 5.5.
Looking Ahead
Using an output cap (max_tokens) for Opus 5 with ‘thinking off’ might cut responses off on Opus 5.5. The guide mentions that ‘thinking’ uses some of that cap even when you don’t see it.
For frontend work without design direction, Opus 5.5 responds well to instructions naming specific patterns to avoid, the guide says. Examples include cream backgrounds and pill-shaped buttons. Vague requests to avoid a generic AI look mostly switch one default for another.
Featured Image: Samuel Boivin/Shutterstock
This article was updated on October 7, 2026 to clarify the guide’s time-budget and chat-prompt guidance. It also reflects a wording change in Anthropic’s documentation.
