Paste your CLAUDE.md. See the per-session token cost, the per-month bill at your plan, and the 5 specific patterns that are eating tokens you do not have to spend.
Updated 2026-05-16. Token estimates use the OpenAI BPE approximation (1 token per ~4 characters of English, ~2 characters of Japanese). Cost projections use post-2026-04-22 Anthropic Pro plan rates and the assumption that CLAUDE.md is read once per turn. The 40% reduction figure comes from a Qiita article with 30,000+ reads and a documented Skills migration of one operator's own configuration.
1. Paste your CLAUDE.md (or SKILL.md, or any always-loaded reference file)
Nothing leaves your browser. The analysis runs locally in JavaScript. Try with your real config; the patterns mean nothing on a placeholder.
2. Which plan are you on?
Used to project the per-month cost of the tokens this file consumes.
3. How many turns per day on average?
A "turn" is one round-trip with Claude. CLAUDE.md is read on every turn, so this multiplies the cost.
120 turns/day
YOUR CLAUDE.MD IS COSTING YOU
Estimated tokens
Per turn (CLAUDE.md is sent every turn)
...
Per-day input tokens from this file
Tokens per turn × turns per day
...
Per-month cost from this file alone
At your plan's rate, assuming uncached. Real cost depends on cache hit ratio.
...
Estimated monthly recovery if you fix the detected patterns
Up to 40% reduction is documented; this projection is the lower bound from detected patterns only.
...
5 PATTERNS THAT EAT TOKENS
These numbers are projections, not measurements. Actual cost depends on your plan's caching behavior, the Sonnet/Opus model mix, and whether CLAUDE.md is loaded as a system prompt (cached) or in the conversation (less reliably cached). The Anthropic 5-minute prompt cache TTL means CLAUDE.md is re-billed at full rate after 5 minutes of inactivity. The pattern detection is heuristic and conservative; for a thorough audit see the linked book. Nothing leaves your browser.