Back to Details

context_mastery

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Context Mastery & Token Optimization Protocol

上下文精通与Token优化协议

1. The "Grep-First" Rule (Token Saver)

1. “Grep优先”规则（Token节省方案）

Problem:
```
read_file
```
on a 2000-line file consumes ~500-1000 tokens instantly.
Solution: Always use
```
grep
```
or
```
read_file_range
```
first to inspect specific functions or imports.
Banned:
```
cat
```
or
```
read_file
```
on entire directories or massive files (e.g.,
```
package-lock.json
```
, giant logs) unless absolutely necessary.

问题：对2000行文件执行
```
read_file
```
会立即消耗约500-1000个Token。
解决方案：始终优先使用
```
grep
```
或
```
read_file_range
```
来检查特定函数或导入内容。
禁止操作：除非绝对必要，否则不要对整个目录或超大文件（如
```
package-lock.json
```
、大型日志文件）执行
```
cat
```
或
```
read_file
```
。

2. Incremental Summarization

2. 增量摘要

Trigger: When the conversation exceeds ~20 turns or you feel the context window filling.
Action:
1. Summarize what has been achieved in the last 10 turns.
2. Write it to
```
scratchpad.md
```
  or
```
active_task.md
```
  .
3. Explicitly state "I am clearing internal context of validated steps" (if tool allows) or just rely on the scratchpad for future lookup.

触发条件：当对话超过约20轮，或你感觉到上下文窗口即将被占满时。
操作步骤：
1. 总结最近10轮对话中已完成的工作。
2. 将总结内容写入
```
scratchpad.md
```
  或
```
active_task.md
```
  。
3. 明确声明“我正在清除已验证步骤的内部上下文”（如果工具支持的话），或者后续直接依赖草稿文件进行查阅。

3. Focused Context Loading

3. 聚焦式上下文加载

Don't: "Read all files in
```
src/
```
to understand the project." (Too expensive).
Do:
1. Read
```
CLAUDE.md
```
  /
```
README.md
```
  .
2. Read directory listing
```
ls -R
```
  .
3. Selectively read only the interfaces (
```
types.ts
```
  ) or entry points (
```
index.ts
```
  ) related to the current task.

避免操作：“读取
```
src/
```
目录下的所有文件以了解项目”（成本过高）。
推荐操作：
1. 读取
```
CLAUDE.md
```
  /
```
README.md
```
  。
2. 读取目录列表
```
ls -R
```
  。
3. 仅选择性地读取与当前任务相关的接口文件（如
```
types.ts
```
  ）或入口文件（如
```
index.ts
```
  ）。

4. Context Efficiency Checklist

4. 上下文效率检查表

Did I use
```
grep
```
instead of reading the whole file?
Did I check if I already have this info in
```
MEMORY.md
```
?
Is my next prompt efficient? (Avoid repeating massive code blocks back to the user).

我是否使用
```
grep
```
替代了读取整个文件？
我是否检查过
```
MEMORY.md
```
中已有相关信息？
我的下一个提示词是否高效？（避免向用户重复大段代码块）