The Cache Discipline Checklist — 7 Moves That Actually Cut Your Claude Bill (After the Viral Tricks Failed Testing)
The companion checklist to the video. A controlled test on Reddit (140 agent runs, 10 real tasks, run twice) found that five of the community's favorite token-saving tricks — caveman prompts, the lazy-engineer persona, an output compressor, a context mode, even Mandarin prompting — cut total tokens exactly zero times reliably; the worst added 68%. What actually works is cache physics, straight from Anthropic's own docs. INSIDE: (1) THE 7 MOVES — /clear between tasks, measure with /usage + /context, subagents downward on Haiku, route models at session start, never compact cold (a 1,263-event dataset put cold compacts at ~15x warm), CLAUDE.md under 200 lines, and cache hygiene as a habit, (2) THE CACHE TABLE — the 8 actions that silently bust your cache (model switch, effort change, fast-mode toggle, MCP connect/disconnect...) vs the 8 that keep it (including /rewind, which lands on an already-cached prefix — and the CLAUDE.md mid-session edit that doesn't even apply), (3) the TTL card — 1 hour on subscription, 5 minutes on API key, subagents always 5 — and the coffee-break rule with the resume-from-summary fix, (4) the myth-bust box with the full test attribution, (5) the Sept 1 Sonnet price note. Every mechanism verified against code.claude.com docs on the day of recording; community numbers attributed to their sources. Independent checklist from Hyperautomation Labs — not affiliated with Anthropic.
Free. No spam. Unsubscribe anytime.