Anthropic Releases Claude Opus 5 Prompting Guide

Anthropic has officially announced a prompt engineering guide optimized for the Claude Opus 5 model, along with a migration guide from Opus 4.8 to Opus 5. The core of this guide is to remove unnecessary instructions used with previous models in order to maximize the model's performance.

Key Changes and Recommendations:

• Prompt optimization: Instructions such as 'verification steps' or 'progress summaries' that were added to compensate for previous models' performance can actually cause excessive verification or redundant narration in Opus 5, so it is recommended to remove them.
• Response length control: Opus 5's responses are longer than before, and the effort parameter only adjusts the model's thinking depth without reducing response length. Length control must be specified directly in the prompt.
• Agentic coding and review: The model shows strength in complex refactoring and feature implementation. For code review, instructing it to 'report only serious issues' can actually reduce the amount reported, so it is more effective to have it report everything and filter separately afterward.
• Efficient resource usage: Since high quality is maintained even at low/medium effort settings, it is recommended to actively use this as a means of controlling cost and latency.
• Vision and long context: A 1 million token context window is the default, and for vision tasks, configuring the model to use tools for iterative analysis is advantageous in terms of both performance and cost.