Context Mastery & Token Optimization Protocol
- The "Grep-First" Rule (Token Saver)
-
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.
- Incremental Summarization
-
Trigger: When the conversation exceeds ~20 turns or you feel the context window filling.
-
Action:
-
Summarize what has been achieved in the last 10 turns.
-
Write it to scratchpad.md or active_task.md .
-
Explicitly state "I am clearing internal context of validated steps" (if tool allows) or just rely on the scratchpad for future lookup.
- Focused Context Loading
-
Don't: "Read all files in src/ to understand the project." (Too expensive).
-
Do:
-
Read CLAUDE.md / README.md .
-
Read directory listing ls -R .
-
Selectively read only the interfaces (types.ts ) or entry points (index.ts ) related to the current task.
- Context Efficiency Checklist
-
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).