/token-audit — Context Token Audit

Purpose: Measure how many tokens your Claude Code configuration consumes before any user task begins. Identify the biggest sources of overhead. Produce a concrete action plan with savings estimates.

When to use:

• You're hitting rate limits before end of day
• Sessions feel slow or context compresses early
• You've added a lot of rules files and want to know the real cost
• After a major config change

What You Will Measure

~/.claude/CLAUDE.md

CLAUDE.md

.claude/rules/*.md

MEMORY.md

Key insight:

.claude/rules/

.md

Step 1 — Run the Measurement

Execute these commands from the project root:

# Component sizes
echo "=== PROJECT CLAUDE.md ===" && wc -c CLAUDE.md 2>/dev/null || echo "none"

echo ""
echo "=== RULES FILES (sorted by size) ===" && find .claude/rules -name "*.md" 2>/dev/null \
  | xargs wc -c 2>/dev/null | sort -rn | head -20

echo ""
echo "=== GLOBAL ~/.claude ===" && ls -la ~/.claude/*.md 2>/dev/null \
  | awk '{print $5, $9}' | sort -rn

Then calculate the full budget:

GLOBAL=$(cat ~/.claude/CLAUDE.md ~/.claude/*.md 2>/dev/null | wc -c)
PROJECT=$(wc -c < CLAUDE.md 2>/dev/null || echo 0)
RULES=$(find .claude/rules -name "*.md" 2>/dev/null | xargs cat 2>/dev/null | wc -c || echo 0)
MEMORY=$(find ~/.claude/projects -name "MEMORY.md" 2>/dev/null \
  | xargs grep -l "$(basename $(pwd))" 2>/dev/null | head -1 \
  | xargs wc -c 2>/dev/null | awk '{print $1}' || echo 0)
TOTAL=$(( GLOBAL + PROJECT + RULES + MEMORY + 30000 ))

echo "Global ~/.claude   : ~$(( GLOBAL / 4 )) tokens ($(( GLOBAL / 1000 ))K chars)"
echo "Project CLAUDE.md  : ~$(( PROJECT / 4 )) tokens"
echo "Rules (auto-loaded): ~$(( RULES / 4 )) tokens"
echo "MEMORY.md          : ~$(( MEMORY / 4 )) tokens"
echo "System prompt      : ~7,500 tokens"
echo "---"
echo "TOTAL fixed context: ~$(( TOTAL / 4 )) tokens"
echo "% of 200K window   : $(( TOTAL / 4 * 100 / 200000 ))%"

Step 2 — Classify Rules Files

For each file in

.claude/rules/

Run this classification prompt:

Read every file in .claude/rules/. For each file, output a table row:

| File | Size (chars) | Class (ALWAYS/SOMETIMES/RARELY/NEVER) | Reasoning (one sentence) |

Sort by size descending within each class.
At the end, calculate: total chars that would leave the fixed context if all
RARELY and NEVER files were excluded. Convert to tokens (÷ 4).

Step 3 — Audit Hook Overhead

Hooks on

PreToolUse

PostToolUse

Check what you have:

# List hooks by event type
python3 - << 'EOF'
import json, os
for path in [os.path.expanduser("~/.claude/settings.json"), ".claude/settings.json"]:
    if not os.path.exists(path): continue
    print(f"\n--- {path} ---")
    data = json.load(open(path))
    for event, hooks in data.get("hooks", {}).items():
        for h in hooks:
            cmd = h.get("command", "?")
            matcher = h.get("matcher", "*")
            print(f"  [{event}] matcher={matcher} → {cmd[:80]}")
EOF

For each

PreToolUse

PostToolUse

/cost

Red flags:

• Hooks that

  cat

  files unconditionally

• git status

  or

  git log

  on every call
• Multi-line echo output for debugging that was never removed
• JSON blobs injected as context

Step 4 — Build the Action Plan

Produce a prioritized table. Rule of thumb: only include actions achievable without external infrastructure (no RAG, no vector databases, no custom MCP servers).

Step 5 — The RAG Question

Lazy-loading via a vector database (RAG) is sometimes pitched as the solution. Assess it honestly before committing:

1. What fixed-context tokens remain after Steps 1-4? (Measure this first.)
2. Is RAG justified? A pgvector + custom MCP setup is a 1-2 week project.
3. Break-even: if you have 10 rules files averaging 3K chars each, classification (30 min) saves as much as RAG would. RAG earns its cost at 50+ rule files where intent-based routing is the only scalable solution.

Output Format

After running the audit, produce this report:

## Token Audit — [PROJECT] — [DATE]

### Budget Summary

| Component | Tokens | % of total |
|-----------|--------|------------|
| Global ~/.claude | X | Y% |
| Project CLAUDE.md | X | Y% |
| Rules (auto-loaded) | X | Y% |
| MEMORY.md | X | Y% |
| System prompt | 7,500 | Y% |
| **TOTAL** | **X** | **100%** |

Context window used before any task: X% of 200K

### Rules Classification

| File | Chars | Class | Action |
|------|-------|-------|--------|
| ... | ... | ALWAYS/SOMETIMES/RARELY | keep/lazy-load/remove |

### Hook Overhead

| Hook | Event | Est. stdout | Calls/session | Total tokens/session |
|------|-------|-------------|---------------|----------------------|
| ... | PreToolUse | X chars | ~Y | ~Z tokens |

### Action Plan

| Action | Savings | Effort | Risk |
|--------|---------|--------|------|
| ... | -X tokens | 30 min | Low |

**Total achievable without infrastructure**: -X tokens → from Y to Z (N% reduction)

### RAG Verdict

[One paragraph: remaining overhead after action plan, whether RAG is justified,
estimated setup cost vs savings.]

Interpreting Results

A 48% reduction is typical after a first-pass audit on a heavily configured project, with no infrastructure changes — just removing the RARELY-used files from auto-load.