Anthropic Publishes Prompting Guidance For Claude Opus 5.5

Anthropic’s prompting guide for Claude Opus 5.5 advises builders to check a number of effort ranges as an alternative of reusing the setting utilized in Claude Opus 5. It additionally suggests chat purposes take away system-prompt strains instructing Claude to think twice earlier than responding.

Opus 5.5 launched Sept. 22. The information says Opus 5 prompts ought to nonetheless work effectively with out edits, and that the corporate’s Opus 5 steering is an inexpensive place to start out.

When the corporate examined this in a chat product, eradicating a think-carefully line made replies begin sooner, with “no clear decline within the high quality of the reply.”

Opus 5.5 Defaults To Medium Effort

Opus 5.5 runs at medium effort by default, one stage beneath the excessive default on Opus 5.

Pondering can’t be turned off on Opus 5.5, and requests that attempt get an error. Opus 5 allowed that at excessive effort or beneath. The information calls effort the primary setting to regulate when balancing high quality towards pace and value.

In Anthropic’s testing, Opus 5.5 at medium effort matches and even surpasses Opus 5 at excessive effort in the case of coding and data work.

To chop pondering, the information says to attempt decrease effort settings earlier than immediate modifications, and save the upper ranges like xhigh and max for duties the place they actually make a distinction in high quality.

Anthropic Suggests Reconsidering “Suppose Fastidiously” In Chat

The information’s motive for suggesting chat apps rethink think-carefully strains is that Opus 5.5 decides for itself how a lot to suppose, with effort as the primary management.

Anthropic’s personal docs embrace that form of line. For Claude Opus 4.7, when effort has to remain low to maintain responses fast, its page on effort suggests including steering comparable to “This job entails multistep reasoning. Think twice earlier than responding.”

Agent Steerage Covers Time Budgets And Pasted Textual content

For agent groups, the information recommends utilizing a time finances primarily based on how lengthy duties may take. With Opus 5.5, the software program helps by monitoring elapsed time.

Anthropic’s exams have proven that small agent teams, when utilizing time alerts, end analysis duties faster than a single agent working with out alerts. Even underneath time budgets, these teams managed to attain reply high quality corresponding to solo brokers. The information emphasizes that the time finances is only a suggestion, and setting a strict timeout might be fairly useful. The mannequin may carry out barely much less completely underneath stress.

For customers who enter textual content from sources like emails, the information suggests utilizing tags with a random ID and including a system observe on learn how to handle tagged textual content. This strategy encourages cautious dealing with and, since tags are plain textual content that may be copied, they supply just one layer of safety towards immediate injection.

5 Claude Opus 5.5 Prompting Suggestions From Anthropic

What Anthropic says to retest, take away, and add when prompting Opus 5.5

• Don’t assume your Claude Opus 5 setting remains to be the proper one for Opus 5.5.

• 2

  Decrease effort earlier than prompting for much less pondering

  Anthropic says effort is the primary setting to regulate when balancing high quality, pace, and value.

• 3

  Contemplate eradicating ‘consider carefully’ in chat prompts

  In Anthropic’s chat-product testing, eradicating that instruction made replies begin sooner.

  Anthropic stated there was ‘no clear decline’ in reply high quality.

• 4

  Use time budgets for predictable duties

  For multiagent workflows with a smart job estimate, Anthropic recommends passing elapsed time towards that finances.

  The finances is advisory, not a tough cease.

• 5

  Tag pasted textual content from exterior sources

  Directions<…>Pasted textual content…>

  Anthropic recommends marking pasted content material and telling Claude learn how to deal with it.

  It describes this as one guardrail towards immediate injection, not a full protection.

Why This Issues

A number of sections of the Opus 5.5 information ask builders to re-check setup carried over from older fashions, comparable to workarounds for studying charts and screenshots, or directions written for operating with pondering off. The Fable 5.1 information, which Roger Montti covered earlier this month, requested builders to revisit formatting guidelines.

An app that by no means units effort runs at medium on Opus 5.5 with none code change.

Trying Forward

Utilizing an output cap (max_tokens) for Opus 5 with ‘pondering off’ may lower responses off on Opus 5.5. The information mentions that ‘pondering’ makes use of a few of that cap even while you don’t see it.

For frontend tasks, the Opus 5.5 information recommends setting clear types to keep away from defaults like cream backgrounds and pill-shaped buttons. Imprecise requests to keep away from a generic AI look usually simply change one default for one more.

Featured Picture: Samuel Boivin/Shutterstock