Claude Code eats more tokens than you expect.

A fresh install drops about 27,000 tokens before you type a single word. On a 200,000-token window that’s 14 % of your capacity gone instantly.

I cloned version 2.2.1 to count what lives in the context:

  • 286 skill names and descriptions – ~21,000 tokens
  • 68 sub-agent descriptions – ~5,000 tokens
  • Instruction files – ~1,000 tokens

Only the skill bodies load on demand; the names and descriptions sit in memory for every session. Every unused skill competes with your actual code and files.

You don’t need the full install. I kept five useful pieces:

  1. AgentShield – audits agent files, hooks, and secrets. It flags malicious skills. Run it once, then verify the version date.
  2. Settings Optimizations – the repo’s .claude/settings.json tweaks cut model costs by 60 %. You trade a smaller default model for lower spend.
  3. Instincts – watches your patterns, writes rules for you, and learns habits without manual input. Requires the hook runtime.
  4. Context Management Tools – advise on token budget. Compact the context after research and before implementation; never do it mid-task.
  5. Selected Agents – the 68 agents cover many languages you’ll never use. Pick the two or three that match your stack.

The goal is to pay only for what you use. Skip whole profiles or prune skills one by one to keep the context window free for your code.

Are you running the full install or just the bits you need? What token count are you starting with?

Source: https://dev.to/aidiveyt/everything-claude-code-costs-27000-tokens-before-you-type-i-kept-five-pieces-52bb

Optional learning community: https://t.me/GyaanSetuAi