context-audit
Audit current context composition and identify optimization opportunities. Use when context window is overloaded, agents are underperforming, or applying the R&D framework to optimize token usage.
When & Why to Use This Skill
This Claude skill provides a comprehensive audit of an agent's context engineering health, identifying optimization opportunities to prevent context window overload. It systematically evaluates memory files, MCP server configurations, and command infrastructure to improve agent performance and token efficiency using the R&D (Reduce and Delegate) framework.
Use Cases
- New Codebase Onboarding: Quickly assess the context composition and memory structure when starting work on a fresh project to ensure a clean working environment.
- Performance Troubleshooting: Identify the root cause of sluggish agent responses or context-related errors by pinpointing bloated memory files or over-provisioned MCP servers.
- Token Usage Optimization: Apply the R&D framework to reduce static context and implement dynamic priming commands for more efficient and cost-effective token consumption.
- Context Health Monitoring: Conduct periodic audits to prevent 'context rot' and ensure the agent remains focused and performant as the project complexity grows.
| name | context-audit |
|---|---|
| description | Audit current context composition and identify optimization opportunities. Use when context window is overloaded, agents are underperforming, or applying the R&D framework to optimize token usage. |
| allowed-tools | Read, Grep, Glob |
Context Audit Skill
Audit a codebase's context engineering health and identify optimization opportunities.
Purpose
A focused agent is a performant agent. This skill helps you understand what's consuming your context window and where to apply the R&D framework.
When to Use
- Starting work on a new codebase
- Agent performance feels sluggish
- Context warnings appearing
- Before optimizing context strategy
- Periodic context health checks
Audit Process
1. Memory File Analysis
Scan for CLAUDE.md and related memory files:
Check:
- Root CLAUDE.md size (target: <2KB)
- Number of imports
- Per-directory CLAUDE.md files
- Total memory file tokens
Score memory health:
| Size | Score | Assessment |
|---|---|---|
| <1KB | Excellent | Minimal and focused |
| 1-2KB | Good | Within target range |
| 2-5KB | Needs Review | Growing, audit content |
| >5KB | Action Required | Bloated, needs R&D |
2. MCP Server Analysis
Check MCP configurations:
Check:
- .mcp.json existence
- Number of MCP servers configured
- Per-server token estimate (2-5% each)
- Active vs unused servers
Score MCP health:
| Servers | Score | Assessment |
|---|---|---|
| 0 | Excellent | No MCP bloat |
| 1-2 | Good | Targeted usage |
| 3-5 | Review | May be over-provisioned |
| >5 | Action Required | Likely consuming 15%+ |
3. Commands Analysis
Review .claude/commands/:
Check:
- Number of commands
- Command complexity (simple vs complex)
- Priming commands present?
- Task-type coverage
Score command health:
| Commands | Score | Assessment |
|---|---|---|
| Has priming | Excellent | Dynamic context loading |
| No priming | Needs Attention | Relying on static memory |
4. Hooks Analysis
Check for context-consuming hooks:
Check:
- Number of hooks
- Hook event types
- Potential context injection
5. Overall Context Score
Calculate overall context engineering score:
| Component | Weight | Max Points |
|---|---|---|
| Memory Files | 30% | 30 |
| MCP Configuration | 25% | 25 |
| Command Infrastructure | 25% | 25 |
| Context Patterns | 20% | 20 |
Output Format
{
"score": 75,
"grade": "B",
"components": {
"memory": {
"score": 20,
"max": 30,
"files_found": ["CLAUDE.md"],
"total_tokens": 1500,
"issues": ["No priming commands detected"]
},
"mcp": {
"score": 25,
"max": 25,
"servers_found": 0,
"estimated_consumption": "0%"
},
"commands": {
"score": 15,
"max": 25,
"count": 5,
"has_priming": false,
"issues": ["Missing /prime command"]
},
"patterns": {
"score": 15,
"max": 20,
"issues": ["No output styles defined"]
}
},
"recommendations": [
"Create /prime command for dynamic context loading",
"Reduce CLAUDE.md size by delegating to priming",
"Consider output styles for token efficiency"
]
}
Grading Scale
| Score | Grade | Status |
|---|---|---|
| 90-100 | A | Elite context engineering |
| 80-89 | B | Good practices, minor optimizations |
| 70-79 | C | Functional, needs attention |
| 60-69 | D | Significant issues |
| <60 | F | Context bloat, major rework needed |
Recommendations Framework
Based on findings, recommend:
For Memory Bloat (Reduce)
- Identify content that can move to priming commands
- Flag outdated or contradictory guidance
- Suggest minimal CLAUDE.md structure
For Missing Infrastructure (Delegate)
- Recommend priming command creation
- Suggest output styles for verbosity control
- Propose agent expert patterns
Cross-References
- @rd-framework.md - Reduce and Delegate strategies
- @context-layers.md - Understanding context composition
- @context-rot-vs-pollution.md - Diagnosing context problems
- @context-priming-patterns.md - Dynamic context loading
Version History
- v1.0.0 (2025-12-26): Initial release
Last Updated
Date: 2025-12-26 Model: claude-opus-4-5-20251101