Anthropic发布Claude Opus 5.5提示工程指南

Anthropic发布了Claude Opus 5.5模型的提示工程指南，详细介绍了该模型相比Claude Opus 5的性能改进与优化方法。1

Claude Opus 5.5在多个方面实现了升级。1该模型生成输出令牌的速度比Claude Opus 5快30%以上，1在多步骤代码工作中，即使在中等努力水平上也能匹配或超过Claude Opus 5的高努力水平性能。1Claude Opus 5.5即使在最低努力设置下，也能比Claude Opus 5在最高努力设置下更准确地读取密集图表中的数值。1

新版本在功能特性上也有重要变化。1Claude Opus 5.5的思考功能始终开启，与允许禁用思考功能的Claude Opus 5不同。1指南建议最大令牌限制设置为128,000，以容纳模型的思考令牌和回复。1同时，Anthropic集成了生物学、网络安全和推理提取等安全分类器。1指南针对代码审查、知识工作、通信和图表识别等能力提供了具体的提示调优建议。1

Anthropic has published a prompting engineering guide for its Claude Opus 5.5 model, detailing behavioral improvements and optimization strategies compared to Claude Opus 5 1. The new model generates output tokens more than 30% faster than its predecessor 1, while introducing several enhanced capabilities including persistent thinking functionality, multi-agent task support, and improved visual input processing 1.

The guide provides specific prompting recommendations across multiple domains, including code review, knowledge work, communication, and chart recognition 1. Notably, Claude Opus 5.5 achieves comparable or superior performance to Claude Opus 5 operating at maximum effort levels even when set to medium effort on multi-step coding tasks 1. In visual analysis, the newer model reads values from dense charts with greater accuracy than Claude Opus 5 at its highest effort setting, regardless of the effort level applied to Opus 5.5 1. The maximum recommended token limit is set at 128,000 to accommodate the model's thinking tokens and responses 1.

A key architectural difference involves the thinking feature: while Claude Opus 5 allows users to disable thinking through specific configuration, Claude Opus 5.5 maintains thinking as permanently enabled 1. The model also incorporates integrated safety classifiers for biology, cybersecurity, and reasoning extraction 1.