In this video I break down 18 token management hacks for Claude Code in 18 minutes, organized into three tiers—from easy wins anyone can do to power-user strategies. The big “light bulb” is how tokens actually work: every time you send a message, Claude rereads the entire conversation from the beginning, so your cost compounds fast. On top of that, Claude Code reloads invisible overhead every turn (cloud.md, MCP tool definitions, system prompts, files), and bloated context doesn’t just cost more—it can also make outputs worse because of “loss in the middle.” From there I get super practical. Tier 1 is about immediate context hygiene: start fresh chats (/clear), disconnect unused MCP servers, batch prompts, use plan mode so Claude doesn’t go down the wrong path, and use /context + /cost (plus a status line) to see what’s actually eating tokens. Tier 2 is about structure: keep cloud.md lean, reference files surgically, compact at ~60% instead of waiting for 95%, avoid command-output bloat, and understand that short breaks can spike costs because caching times out. Tier 3 is where you start playing offense: pick the right model (Sonnet default, Haiku for sub-agents/simple tasks, Opus only when needed), be careful with sub-agent workflows (they’re expensive), and schedule heavy work for off-peak hours. My core takeaway: most people don’t need a bigger plan—they need better context management.