Claude Token Hygiene — Security skill for Claude Code

Claude Code Token Hygiene

**Your Claude Code conversations are wasting 30-50% of tokens on system context you don't need.**

Every conversation loads CLAUDE.md, MEMORY.md, skill descriptions, MCP tool listings, and hooks — before you type a single word. On a typical power-user setup, that's **15,000-35,000 tokens of overhead per conversation**.

This skill audits that overhead and tells you exactly where to cut.

---

The Problem

You type: "fix the bug in auth.ts"

What Claude actually processes:
  ├── CLAUDE.md instructions        ~8,000 tokens
  ├── MEMORY.md auto-memory         ~5,000 tokens
  ├── 89 skill descriptions         ~6,000 tokens
  ├── 7 MCP servers (220 tools)     ~3,300 tokens
  ├── Hooks & system reminders      ~2,000 tokens
  ├── Your message                      ~20 tokens
  └── Total                        ~24,320 tokens  ← before Claude even thinks

On Opus at $15/M output tokens, that overhead costs **real money** on every conversation.

The Solution

A monthly audit that measures your system context, flags what's bloated, and recommends specific cuts:

Token Hygiene Report — 2026-03-05

┌──────────────┬───────────────┬────────┬─────────────┐
│ Metric       │ Value         │ Status │ Tokens Est. │
├──────────────┼───────────────┼────────┼─────────────┤
│ CLAUDE.md    │ 123 lines/6KB │   ✅   │ ~1,600      │
│ MEMORY.md    │ 90 lines/5KB  │   ✅   │ ~1,350      │
│ Skills       │ 82            │   🟡   │ ~5,330      │
│ MCP servers  │ 5             │   ✅   │ ~1,125      │
│ Stale logs   │ 0             │   ✅   │ —           │
│ Total        │               │        │ ~9,405      │
└──────────────┴───────────────┴────────┴─────────────┘

↓ Down from ~30,000 tokens (previous audit)

Real Results

I ran this on my own setup and reduced overhead from **~30,000 to ~15,000 tokens per conversation** — a 50% reduction:

What I did	Savings	Time
Compressed CLAUDE.md (713→123 lines)	~6,7