Token Hygiene — Context Overhead Audit

Monthly audit of system files that inflate every Claude Code conversation's token cost. Reports metrics, flags issues, and recommends actions. Does NOT modify files automatically.

Why This Matters

Every Claude Code conversation loads your system context before you type anything:

Source	What it is	Typical size
CLAUDE.md	Project instructions	2,000-10,000 tokens
MEMORY.md	Auto-memory (if enabled)	1,000-6,000 tokens
Skill descriptions	One line per installed skill	50-80 tokens each
MCP tool listings	Every registered tool	~15 tokens each
Hooks & plugins	System reminders	500-2,500 tokens

A typical setup can burn 15,000-35,000 tokens per conversation just on system context — before you ask a single question. This skill measures that overhead and tells you where to cut.

Thresholds

Check	What it measures	Green	Yellow	Red
CLAUDE.md size	Bytes of project instructions	<10KB	10-20KB	>20KB
MEMORY.md lines	Lines in auto-memory file	<150	150-200	>200
Skills count	Folders in ~/.claude/skills/	<80	80-100	>100
MCP servers	Servers in .mcp.json	<=5	6-7	>7
Stale daily logs	memory/*.md files older than 30 days	0-5	6-15	>15

Workflow

Step 1: Measure All Sources

# CLAUDE.md
wc -l CLAUDE.md && wc -c CLAUDE.md

# MEMORY.md (auto-memory location varies by project)
wc -l ~/.claude/projects/*/memory/MEMORY.md
wc -c ~/.claude/projects/*/memory/MEMORY.md

# Skills count
ls -d ~/.claude/skills/*/ 2>/dev/null | wc -l

# MCP servers
grep -c '"command"' .mcp.json 2>/dev/null

# Stale daily logs (if you use daily memory logs)
find memory/ -name "20*.md" -mtime +30 2>/dev/null | wc -l

Step 2: Estimate Token Impact

Source	Formula
CLAUDE.md	bytes × 0.28
MEMORY.md	bytes × 0.33
Skills	count × 65
MCP tools	servers × avg_tools × 15

Step 3: Identify Issues

For each metric in Yellow or Red:

CLAUDE.md too large:

• Look for reference sections (tables, examples, long command lists)
• Move them to a separate file and add a one-line pointer
• Remove stale content (tools you've uninstalled, old client lists)

MEMORY.md too large:

• Extract detailed content into topic files (e.g., memory/client-details.md)
• Keep only operational rules and pointers in MEMORY.md
• Remove entries older than 2 months that aren't actively referenced

Too many skills:

• Check for duplicates (same skill from different sources)
• Uninstall skills you haven't used in 30+ days
• Each removed skill saves ~65 tokens per conversation

Too many MCP servers:

• Remove servers with expired credentials
• Disable servers you use rarely (re-enable when needed)
• Check for duplicates (e.g., plugin playwright + project playwright)

Step 4: Generate Report

The report uses a traffic-light dashboard:

# Token Hygiene Report — [DATE]

## Dashboard

| Metric | Value | Status | Tokens Est. |
|--------|-------|--------|-------------|
| CLAUDE.md | X lines / Y KB | [status] | ~N tokens |
| MEMORY.md | X lines / Y KB | [status] | ~N tokens |
| Skills installed | X | [status] | ~N tokens |
| MCP servers | X | [status] | ~N tokens |
| Stale logs | X files | [status] | — |
| **Total overhead** | | | **~N tokens** |

## Issues Found
[Details per issue with recommended action and estimated savings]

## Recommended Actions (Priority Order)
1. [Action] — saves ~N tokens
2. [Action] — saves ~N tokens

Step 5: Track Progress

Save results to compare month-over-month:

{
  "last_audit": "2026-03-05",
  "metrics": {
    "claude_md_bytes": 5669,
    "memory_md_lines": 90,
    "skills_count": 82,
    "mcp_servers": 5,
    "stale_logs": 0,
    "estimated_overhead_tokens": 17000
  }
}

Common Wins

Based on real-world optimization (40-50% reduction achieved):

Action	Typical Savings	Effort
Move CLAUDE.md reference sections to separate file	3,000-6,000 tokens	30 min
Compress MEMORY.md, extract to topic files	2,000-4,000 tokens	20 min
Remove duplicate skills	200-500 tokens	5 min
Remove unused MCP servers	500-2,500 tokens	5 min
Disable cloud MCP integrations (Supabase, etc.)	500-1,500 tokens	2 min
Archive old daily logs	Reduces file scan time	5 min

Notes

• This skill is read-only. It never modifies files automatically.
• You review the report and decide which actions to take.
• Run monthly, or whenever conversations feel slower than usual.