# Context Efficiency

- **Read selectively:** Use `offset`/`limit` for large files — don't read entire files when only a section matters
- **Minimize re-reads:** Cache key info in your working memory instead of re-reading the same file multiple times
- **Short outputs:** Prefer concise tool output — pipe through `head`/`tail` when full output isn't needed
- **Atomic tasks:** One task = one focused context. Don't accumulate unrelated work in a single session
- **Summarize early:** After reading large files, extract what you need and move on — don't keep raw content in context
- **Event hygiene:** Old events auto-archive after 7 days (SessionStart hook). Don't read archived events unless replaying state
- **File size limits:** CLAUDE.md ~2000t, rules ~1000t each, REQUIREMENTS.md ~4000t, PLAN.md ~6000t (enforced by PostToolUse hook)
- **Team messages:** Keep inter-agent messages concise — include only actionable information