token-optimize

Measures and reduces tokens consumed by agent descriptions, skill descriptions, CLAUDE.md files, and memory files that load into every Claude Code conversation.

Why this matters

Every conversation starts with a fixed token budget consumed by system prompt contents. Agent descriptions, skill descriptions, CLAUDE.md, and MEMORY.md all load unconditionally. Reducing these frees context for actual work. Korean text costs 2-4x more tokens than equivalent English for the same semantic content.

Workflow

Phase 1: Measure

Always run measurement first. Use the bundled script:

# Project only (default)
uv run python <skill-dir>/scripts/measure.py --project-dir <project-root>