Unverified85% confidenceFactTime unknown
Anthropic建议CLAUDE.md文件保持简洁，因为过长的内容会导致Claude开始忽略部分信息，这与Transformer架构中的注意力稀释现象有关
1
Sources
85%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
Claude Code官方推荐的7条最佳实践指南
bilibili云桥网络
Related Entities
Related Claims
VerifiedTransformer架构的注意力机制在处理超长序列时，对早期内容的关注度会显著下降，导致模型遗忘前文设定68% similarUnverified文档分块策略中过长则检索精度下降，过短则丢失上下文66% similarUnverified给Claude Opus写过于详细的步骤反而会限制其推理空间，而给DeepSeek写过于抽象的目标则会导致执行偏差65% similarUnverifiedCLAUDE.md 每个文件建议控制在 200 行以内，超过这个长度 Claude 的遵从度会明显下降64% similarVerified文档切分策略中切分过细会丢失上下文，切分过粗会引入噪声64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/35930API
curl https://kongchang.com/api/v1/knowledge/claims/35930MCP
get_claim(id=35930)