Jannis Rothe - How to Cut Your Claude Token Usage by 60%+

Original author：Claude Community Australia

Original source：https://www.youtube.com/embed/OuD4bQxYbzw

Jannis Rothe is a data analytics specialist at DMN Digital who taught himself to get far more out of Claude Code on fewer tokens. In this talk he runs through practical savings: /clear and /compact, avoiding MCP servers, building skills, using cheaper models, converting PDFs to Markdown, plus the Caveman and RTK (Rust Token Killer) skills and his own quota status-line and handoff skill. A grounded, security-aware guide to spending fewer tokens while getting better results. Recorded at Claude for Everyone - Sydney, 23 June 2026. Organised by the Claude Community Australia - claudecommunity.au Timestamps: 0:00 Intro 0:03 The bold claim: cut token usage by 60%+ 0:43 Show of hands: who uses what part of Claude 1:23 Background: a non-engineer who loves Claude Code 2:04 What a token is, and why it costs money and quality 3:21 /clear vs /compact and using fresh sessions 4:01 Architecture: avoid MCP servers where you can 5:22 Why MCP servers always load and bloat context 6:02 Build skills for repetitive tasks 6:42 Not everything needs to live in Claude 8:02 Convert PDFs to Markdown to save tokens 8:43 Use cheaper and local models for easier tasks 10:04 Better context up front and interview skills 11:25 Plug-and-play tools that reduce tokens on autopilot 12:06 Security warning: vet skills before you download 14:46 The Caveman skill: ultra-compact response mode 16:07 The pasta example: full vs condensed output 18:08 RTK (Rust Token Killer): a token-stripping hook 19:28 Quota status-line: track 5-hour and 7-day usage 21:33 The handoff skill paired with /clear 23:35 Q: Is GitHub the right place to get these? 26:13 Q: How did you start reducing tokens? 27:33 Q: What if I need an MCP for an API? 28:54 Q: How does Caveman affect actual code output? 30:15 Q: Can you stack RTK and Caveman together?