Anthropic releases official Claude Opus 5 prompting guide: key recommendations

Company Anthropic Published an official guide to compiling prompts for its language model Claude Opus 5. In the instructions, the developers focused on the correct configuration of requests, which avoids excessive consumption of tokens.

Requests for brevity and control of subagents

The main difference between Opus 5 and previous versions is the tendency to comment in detail on their own actions and automatically start subagents to perform tasks.

To prevent the model from spending the entire token limit on reasoning, Anthropic recommends:

• Add direct instructions on the brevity and conciseness of the answer.
• Limit in-depth process details when there is no direct need.

Exclusion of unnecessary instructions

The developers separately note that requests such as “recheck again” or “make sure the answer is correct” should be removed from the prompts.

In Claude Opus 5, the self-check algorithm is built in by default – the model performs the verification of the results automatically during generation, so additional instructions in the query text become redundant.

Where to read the manual

The full version of the prompt engineering guide is available in Official documentation of Anthropic.