context-optimizer（上下文优化器）

Context Optimizer

Manage your context window and token budget effectively.

Quick Diagnosis

1. Run /context to check current usage
2. If > 70% → compact now before it degrades
3. If > 90% → you're in the "dumb zone", compact immediately

Optimization Strategies

Immediate

Configuration

Set proactive auto-compaction:

{
  "env": {
    "CLAUDE_AUTOCOMPACT_PCT_OVERRIDE": "50"
  }
}

MCP Audit

Keep <10 MCPs enabled, <80 tools total. Each MCP adds overhead to every request.

/mcp          # List active servers
# Disable what you're not using

Prompt Engineering for Token Efficiency

• Scope your prompts: "In src/auth/, fix the login bug"
• Provide constraints: "Don't modify the middleware"
• Give acceptance criteria: "Should return 429 after 5 attempts"
• Avoid vague prompts: "Fix the code" (forces Claude to read everything)

Subagent Delegation

Heavy operations that generate lots of output should go to subagents:

• Test suite output → subagent
• Large file exploration → subagent
• Documentation generation → subagent
• Log analysis → subagent

The main session stays clean while subagents handle the volume.

Context Budget Planning

Token Efficiency

Output Reduction (40-60% savings)

• No sycophantic openers ("Sure!", "Great question!")
• No closing fluff ("Let me know if you need anything!")
• No prompt restatement before answering
• Code first, explanation only if non-obvious
• Structured output (tables, bullets) over prose
• ASCII only: -- not em dashes, straight quotes not smart quotes

Behavioral Efficiency

• One-pass coding: complete solution, test once, stop if green
• Read before write: never modify unread files
• No re-reads: don't re-read unchanged files
• Tool-call budgets: 20 (quick fix) to 80 (large feature)
• Never iterate more than twice on the same failure

Task Profiles

Switch response style based on context:

• Coding: code first, minimal explanation, simplest solution
• Agent/Pipeline: structured output only, no prose, parseable
• Analysis: finding first, tables over paragraphs, sourced numbers

CLAUDE.md Optimization

• Root CLAUDE.md: < 60 lines ideal, < 150 max
• Move package-specific info to package-level CLAUDE.md
• Move personal preferences to CLAUDE.local.md
• Remove obvious or rapidly-changing information

When Context Is Degraded

Signs:

• Claude repeats itself or forgets earlier context
• Responses become generic or lose project-specific knowledge
• Tool calls start failing for reasons that worked earlier

Fix:

1. Manual /compact
2. If still bad: new session with /resume
3. For recurring issues: reduce CLAUDE.md size, disable MCPs