Claude Code 实践心得：飘才是最大的敌人
英文摘要
The author now writes code entirely with Claude Code. After experimenting with multi-agent frameworks and tools like superpowers, they find that these systems produce visually correct UIs but messy internal code, often taking hours. Programmers dislike the loss of transparency, as real-world debugging falls back on them. The core problem is 'drift': AI models lose context like a game of telephone, and multi-agent pipelines amplify errors far beyond original intent.
中文摘要
作者如今已完全使用Claude Code编写代码。尝试过多智能体框架和类似superpowers的工具后，发现这些系统往往生成看似正确的UI，但内部代码一团混乱，且耗时数小时。程序员不喜欢这种不透明感，因为实际调试仍需自己动手。核心问题是“飘”：模型会像传声筒游戏一样失去上下文，多智能体上下游关系更会放大错误，使最终输出与原始意图严重偏离。
关键要点
The author has fully adopted Claude Code for all coding tasks.
作者已全面采用Claude Code进行编程。
Multi-agent coding tools often produce surface-level correct results but deliver messy internal code after long waits.
多智能体编程工具常生成表面正确的结果，但经过长时间等待后给出的内部代码混乱不堪。
Programmers value transparency, as debugging opaque agent-generated code ultimately falls on them.
程序员重视透明性，因为调试不透明的智能体生成代码最终仍需自己动手。
The fundamental issue is drift: models deviate from instructions, and multi-agent chains compound these errors, making results unreliable.
根本问题是“飘”：模型会偏离指令，多智能体链条会累积错误，使结果不可靠。