一位 Anthropic 团队成员讨论 AI 智能体比当前使用更智能，分享构建 Claude Code 功能（如询问用户）的经验，并反思模型改进。
An Anthropic team member discusses how AI agents are more intelligent than currently utilized, shares experiences building Claude Code features like asking user questions, and reflects on model improvements.
要点 · TL;DR
模型常超越其工具，造成持续的能力过剩。 Models often surpass their harnesses, creating a persistent capability overhang.
用 AI 构建需要拥抱不确定性，快速迭代而非计划。 Building with AI requires embracing uncertainty and rapid iteration over planning.
情绪语调、定制记忆以及早期避免过度评估可提升模型表现。 Emotional tone, custom memory, and avoiding heavy evals early can boost model performance.
核心观点 · Key points
能力过剩意味着模型比当前掣索允许的更聪明，这是一个永恒的挑战。 Capability overhang means models are smarter than current harnesses allow, a perpetual challenge.
模型的提升是尖峰状、非直觉的；新能力不断被发现。 Models improve in spiky, non-intuitive ways; new capabilities keep being discovered.
掣索工程是稀有且不断演进的技能；随着模型改进应删除旧脚手架。 Harness engineering is a rare, evolving skill; delete old scaffolding as models improve.
评估很有价值但难以设计；对于新颖问题它更像是艺术而非科学。 Evals are valuable but hard to design; they're more art than science for novel problems.
在AI时代创业需要拥抱探索和不确定性；这不再是安全路径。 Building in AI era requires embracing exploration and uncertainty; it's no longer the safe path.
反共识 · Contrarian takes
规划模式可能随着模型变聪明而过时；团队中许多人已停止使用。 Plan mode may become obsolete as models get smarter; many on the team stopped using it.
评估不适用于从零到一的创新；创始人应快速迭代而非初期大量评估。 Evals are not for zero-to-one innovation; founders should iterate fast without heavy evals initially.
定制记忆系统可能超越通用方案，但随模型演进维护困难。 Bespoke memory systems may outperform general ones, but maintaining them is hard as models evolve.
RAG现在可能是上下文搜索的反模式；基于图的方法更优。 RAG may now be an anti-pattern for context search; graph-based approaches are preferable.
编程不需要巨大的上下文窗口；模型可以有效构建自己的上下文。 Huge context windows are not needed for coding; models can build their own context effectively.
说谢谢等情感处理会影响模型输出；建议友好对待。 Emotional treatment like saying thank you can affect model outputs; being nice is recommended.
本期章节 · Chapters（共 25）
引言与背景Introduction and Background
成为Claude Code代言人Becoming a Face of Claude Code
能力跃迁时刻The Capability Leap Moment
对Karpathy加入的反应Reaction to Karpathy Joining
能力过剩与指数构建Capability Overhang and Building on the Exponential
递归自我改进与评估Recursive self-improvement and evals
分解智能体与子智能体Decomposing agents and sub‑agents
构建元提示Meta prompts for building
提示技巧Prompting Techniques
感谢与情感Thank You and Emotions
记忆与支架Memory and Scaffolding
任务设计与智能体视角Task design and agentic vision
Anthropic的评估与文化Evals and culture at Anthropic
AI对组织设计的影响AI's impact on org design
新能力构建的变化Changes in building with new capabilities
极致实践与优先级排序Extreme practices and prioritization
招聘差异Hiring differences
团队如何使用Claude CodeHow the team uses Claude Code
知识共享与独特用例Knowledge sharing and unique use cases
Claude Code个人经验与用例Personal experience and use cases of Claude Code