Claude Code 重度用户经常遇到上下文溢出导致模型变笨的问题，这个 60% 规则和手动压缩技巧能帮你精准控制保留内容，避免自动压缩丢失关键信息，建议试试。
开发者 Santiago 分享了一个 Claude Code 的使用技巧：当上下文利用率超过 60% 时，模型性能会下降，且自动压缩可能丢失重要信息。他建议通过 /statusline 命令实时监控上下文百分比，并在超过阈值时手动运行 /compact 命令，附带提示词指定保留内容（如项目描述和变更指令）。这样能更精准地控制上下文压缩，避免关键信息丢失。
The 60% rule to compact your context: Don't let y…
The 60% rule to compact your context:
Don't let your sessions go above 60% in context utilization. After that, your model will start feeling drunk and you'll risk Claude Code to trigger auto-compact for you.
Auto-compact is not good if you want control on what context stays and what goes.
Instead:
• Run /statusline to permanently display context % • When > 60%, run /compact with instructions on what to keep • For even more control, dump session context to a file and edit it
Many people don't know this:
You can run /compact with a prompt explaining what you want to keep. For example:
"/compact Keep the project description and change instructions"
This will anchor Claude Code's summary around those topics and will compress everything else more aggressively.