Anthropic在发布Claude Opus 5时删掉了超过80%的系统提示词，但模型表现基本保持不变。[1][2]这一看似矛盾的举措背后，反映的是AI模型能力提升后提示词工程的重大转向。工程师通过对比实际请求数据发现，Opus 4.7版本的系统提示词约有15225个字符，4.8版本精简至4467个字符，但Opus 5版本随后又增加至7694个字符。[1]被删除的内容主要涉及为旧模型设计的六个方向的判断规则：包括让模型自主判断、设计接口而非给出事例、按需加载规则、简化重复说明、由模型自动记忆以及只提供简单规格。[1]
这些被删掉的提示词曾是让Claude Code成为强大Agent的关键。[1]然而，随着模型能力的演进，这些规则开始对产品造成拖累，最终被精简。与此同时，Anthropic在新版本中增加了约72%的新内容，主要包括两段新的行为限制规则。[1]其中"Delivering work"要求模型在原定范围内交付工作，不能悄悄改变任务；"Corrections"则规定只解释真正影响结果的错误，避免反复批评。[1]Claude Code之父Boris Cherny在Y Combinator的访谈中介绍了Opus 5的突破性能力：模型可连续运行超过15天完成长期复杂任务，提示注入攻击基本失效，Bun团队甚至用Claude Code将10万行代码库从Zig重写为Rust，整个过程耗时11天。[2]这些进展表明，通过精简过时规则并引入新的智能限制，AI编程助手的实际能力获得了质的提升。
Anthropic's decision to dramatically reduce Claude Opus 5's system prompt has sparked detailed technical analysis revealing a more nuanced story than initial claims suggest. While the company stated that removing over 80% of the prompt text produced no degradation in model performance, engineer Chen Cheng's examination of actual API requests uncovered the full picture: Opus 4.8 compressed the prompt from approximately 15,225 characters to 4,467 characters, yet Opus 5 subsequently expanded it by 72% to reach 7,694 characters.[1]
The removed content primarily consisted of decision-making rules designed for earlier model iterations, including instructions for the model to make independent judgments, design interfaces rather than provide examples, defer loading information until needed, avoid repetition, rely on automatic memory, and employ simplified specifications.[1] These deletions stripped away what had made Claude Code function as a powerful autonomous agent. However, as the underlying model evolved, these legacy rules began to undermine product performance rather than enhance it.
The new additions to Opus 5's prompt focus on two primary behavioral directives: one governing work delivery by requiring adherence to original scope without unilateral task modifications, and another addressing error correction by limiting explanations to issues that genuinely affect outcomes rather than offering repeated criticism.[1]
Claude Code's creator Boris Cherny revealed in a Y Combinator conversation that Opus 5 can now run continuously for over 15 days to complete extended tasks, with prompt injection attacks rendered essentially ineffective through a three-layer defense mechanism combining alignment training, injection classification, and Auto Mode.[2] The model has demonstrated capability to rewrite 100,000 lines of code from Zig to Rust—a process executed in 11 days by the Bun team using Claude Code.[2] Dynamic workflows now allow Claude Code to orchestrate thousands of parallel agents for complex task execution, with the system launching hundreds to thousands of agents daily while running 20 to 30 routine maintenance operations per day.[2]
