Claude Code Token Guide
by Walid Boulanouar / AY Automate
practical guide
Save 75% on Claude Code tokens
8 techniques ranked by impact. One cache bug you must fix today.
75%
Caveman reduction
10-20x
Cache bug waste
30x
CLI over MCP savings
/month
/month
8
Techniques ranked by savings — descending staircase
Fix the cache bugcritical bug
Anthropic cache invalidation wastes 10-20x tokens silently. github.com/Rangizingo/cc-cache-fix
Caveman mode
Strip verbose system prompts to essentials. github.com/JuliusBrussee/caveman
Ollama local routing
Run exploration tasks locally. Zero API cost. youtu.be/O2k_qwZA8HU
Context Mode
Load files on demand, not everything upfront. mksg.lu/blog/context-mode
Multi-model routing
Haiku for exploration, Sonnet for daily, Opus for complex only.
CLI over MCP tools
Native CLI returns compact results. MCP tools return bloated JSON.
Scoped tool permissions
Deny unused tools. Fewer tools = smaller system prompt.
tldr over man pages
Man pages dump thousands of tokens. tldr gives the essentials.
Session resume, not restart
Resume keeps cached context. Restart re-reads everything from scratch.
The rate limit drama is real. But half the problem is token waste. Fix the waste first.
Quick-start commands — click to copy