Claude Code创始人Boris Cherny近日在播客中分享了Opus 5模型的应用经验与系统优化策略。[1]他建议开发者每隔六个月删除claude.md、Skills和Hooks等系统提示词与工具集合，由模型根据自身能力重新决定所需功能。[1]这一策略的效果显著——Claude Code在Opus 5发布后删除了超过80%的系统提示词。[1]
Opus 5在ARC-AGI-3基准测试中达到30%的成绩，可连续运行数月而无需额外的技术支撑。[1]Boris强调，使用AI编程工具的方法论已从"依赖理论的学问"转变为"不断试错的经验科学"，需要通过实际实验而非预先设计来优化模型使用。[1]Dynamic Workflows框架能够协调数千甚至数万个智能体完成复杂任务，其中Anthropic已通过Routines实现代码库的自动化维护，每天运行数百至数千个智能体。[1]
现实应用中已涌现多个成功案例。[1]Bun团队用Opus 5花费11天时间将整个代码库从Zig重写为Rust，并已投入生产环境。[1]Boris启动的Swift重写实验运行已超过两周，其背后可能调用数千至数万个智能体。[1]
Boris Cherny, creator of Claude Code, recently discussed best practices for leveraging the Opus 5 model in a podcast, revealing an unconventional approach to prompt engineering and agent orchestration.[1]
Cherny recommends deleting core system components every six months, including claude.md, Skills, and Hooks, allowing the model to independently determine what capabilities it requires.[1] This strategy reflects a broader shift in how developers interact with advanced AI models—moving away from theoretical design principles toward empirical experimentation.[1] Following Opus 5's release, Claude Code removed more than 80% of its original system prompts, demonstrating Cherny's confidence in the model's self-directed capability evolution.[1]
The Opus 5 model has demonstrated remarkable performance, achieving 30% accuracy on ARC-AGI-3 benchmarks and operating continuously for months without additional scaffolding support.[1] Cherny shared concrete examples of this capability in action: the Bun team successfully rewrote an entire codebase from Zig to Rust in 11 days using Opus 5, with the resulting code already deployed to production.[1] Cherny's own Swift rewrite experiment has been running for over two weeks, likely orchestrating thousands to tens of thousands of autonomous agents working in parallel.[1]
Dynamic Workflows technology enables coordination of thousands or even tens of thousands of agents on complex tasks, according to Cherny.[1] At Anthropic, the Routines feature automates codebase maintenance by running hundreds to thousands of agents daily.[1] Cherny's advocacy suggests that successful AI-powered programming increasingly depends on experimentation and observation rather than predetermined theoretical frameworks.[1]