token-optimize

Token Optimize

Overview

Apply targeted compression to reduce token usage in specific content — files, conversation history, tool outputs, or prompts — while preserving semantic meaning.

Compression Techniques

Technique	When to Use	Est. Savings
Remove redundancy	Repeated explanations, duplicate imports	10–30%
Summarize	Conversation history, verbose docs	50–80%
Use reference	Inline code that could be "see line X"	20–50%
Compress patterns	Repeated structures, long lists	20–40%
Truncate	File contents where partial read suffices	30–70%
Deduplicate	Duplicate tool outputs, repeated reads	20–60%

Process

1. Confirm target content. If called from token-analyze, the target section is already identified.

2. Estimate current token count (1 token ≈ 4 chars).

3. Select techniques by content type:

   • Conversation history → summarize (preserve decisions, discard exploration)
   • File contents → truncate + use reference
   • Tool results → remove-redundancy + summarize
   • Prompts/rules → remove-redundancy + compress-patterns

4. Produce compressed version. Show a diff summary (key additions/removals), not the full diff unless asked.

5. Estimate savings. Show: [original] → [compressed] tokens ([X]% reduction).

6. Request confirmation before applying to any live context. Never modify CLAUDE.md or SKILL.md files without explicit user approval.

Common Mistakes

• Over-compressing: Content needed shortly will have to be re-read, costing tokens. Ask about future needs before aggressive truncation.
• Applying without confirmation: Always show the plan first.
• Not tracking removals: Note what was omitted so the user can recover it if needed.