| name | demerzel-context-budget |
| description | Audit context window consumption across skills, rules, MCP tools, CLAUDE.md, and memory. Surface bloat and recommend optimizations. Inspired by ECC context-budget. |
Demerzel Context Budget — Token Overhead Audit
Analyze token consumption across every loaded component and surface actionable optimizations.
Usage
/demerzel context-budget — full audit
/demerzel context-budget --quick — summary only
Why This Matters
With 36+ skills, 22 policies, .claude/rules/, CLAUDE.md, memory files, and 200+ MCP tool schemas loaded, sessions can burn 50K+ tokens before the user says anything. This skill measures the actual overhead and identifies savings.
How It Works
Phase 1: Inventory
Scan all component directories and estimate token consumption (words x 1.3):
Skills (.claude/skills/*/SKILL.md)
- Count tokens per SKILL.md
- Flag: files > 400 lines (heavy skills)
- Note: skills are loaded on-demand, but the skill LIST is always in context
Rules (.claude/rules/*.md)
- Count tokens per rule file
- These are ALWAYS loaded — every rule costs tokens every message
- Flag: rules > 100 lines
CLAUDE.md chain
- Project CLAUDE.md + any parent CLAUDE.md files
- Always loaded — direct context cost
- Flag: combined > 300 lines
Memory (~/.claude/projects/.../memory/MEMORY.md)
- Always loaded (first 200 lines)
- Referenced memory files loaded on access
- Flag: MEMORY.md index > 150 lines
MCP Tool Schemas
- Count configured MCP servers and total tool count
- Estimate ~500 tokens per tool for schema overhead
- Flag: servers with > 20 tools
Personas (personas/*.persona.yaml)
- Not loaded by default, but referenced by skills
- Count for awareness
Phase 2: Classify
| Bucket | Criteria | Action |
|---|
| Always loaded | Rules, CLAUDE.md, MEMORY.md index, skill list | Optimize aggressively |
| On-demand | Skills (loaded when invoked), memory files | Optimize if large |
| Background | MCP schemas (loaded when tool used) | Monitor tool count |
| Never loaded | Personas, schemas, policies (not in context) | No action needed |
Phase 3: Detect Issues
- Rule bloat — rules that duplicate CLAUDE.md content
- Memory bloat — MEMORY.md index too long, stale memories
- Skill bloat — SKILL.md files that are unnecessarily verbose
- MCP over-subscription — too many servers registered
- CLAUDE.md sprawl — verbose instructions that should be rules or skills
Phase 4: Report
Context Budget Report
═══════════════════════════════════════
Total estimated overhead: ~XX,XXX tokens
Context model: Claude Opus 4.6 (1M window)
Effective available: ~XXX,XXX tokens (XX%)
Component Breakdown:
┌─────────────────┬────────┬───────────┐
│ Component │ Count │ Tokens │
├─────────────────┼────────┼───────────┤
│ Rules │ N │ ~X,XXX │
│ CLAUDE.md │ N │ ~X,XXX │
│ Memory index │ 1 │ ~X,XXX │
│ Skill list │ N │ ~X,XXX │
│ MCP schemas │ N │ ~XX,XXX │
└─────────────────┴────────┴───────────┘
Issues Found (N):
[ranked by token savings potential]
Top Optimizations:
1. [action] → save ~X,XXX tokens
2. [action] → save ~X,XXX tokens
3. [action] → save ~X,XXX tokens
Potential savings: ~XX,XXX tokens (XX%)
Phase 5: Act (optional)
If --fix flag is passed:
- Trim verbose rules
- Archive stale memory entries
- Suggest CLAUDE.md consolidation
- Report changes made
Governance
- Article 8 (Observability) — context consumption is a health metric
- Article 4 (Proportionality) — loaded components should match session needs
- Staleness detection — stale memory entries waste context
- This skill should be run at the start of long sessions
Credit
Inspired by Everything Claude Code context-budget skill. Adapted for Demerzel's governance framework.
