Harness Audit

Harness Health Audit

Diagnoses the overall health of a Claude Code harness across 8 dimensions.

The 8 dimensions

1. CLAUDE.md quality (0-3)

• 0: none
• 1: exists but basic
• 2: project context + rules included
• 3: detailed architecture + patterns + anti-patterns

2. Skills coverage (0-3)

• 0: no skills
• 1: 1-5
• 2: 6-15 + trigger/anti-trigger definitions
• 3: 16+ + domain classification + eval skill included

The audit target is not a fixed skill list — scan the full set of skills actually present in the skills/ directory.

3. Agents architecture (0-3)

• 0: no agents
• 1: 1-3 (basic)
• 2: 4-8 + clear role separation
• 3: 9+ + separate evaluator + role boundary matrix

4. Hooks automation (0-3)

• 0: no hooks
• 1: basic (prettier, lint)
• 2: quality gates + security scan
• 3: full pipeline (coverage + security + e2e + eval)

5. Rules structure (0-3)

• 0: no rules
• 1: 1-2 general rules
• 2: file-pattern rules (*.ts, *.py, etc.)
• 3: per-domain + security + performance + testing rules

6. MCP servers (0-3)

• 0: no MCP
• 1: 1-2 basic
• 2: search + docs + UI tools
• 3: search + docs + code analysis (code-review-graph, etc.)

7. Eval pipeline (0-3)

• 0: no eval
• 1: manual verify only
• 2: automated quality gates + scoring
• 3: independent evaluator + pass@k + AI slop detection

8. Team/multi-agent (0-3)

• 0: solo use
• 1: basic subagents
• 2: role-based team + communication protocol
• 3: TTH-grade silos + Ralph Loop + backpressure

Procedure

1. Scan the entire ~/.claude/ directory
2. Score each dimension (0-3)
3. Total (0-24) + grade

Grades

Output format

📊 Harness health: [Grade] ([N]/24)

| Dimension | Score | Status |
|-----------|-------|--------|
| CLAUDE.md | [N]/3 | [🟢/🟡/🔴] |
| Skills | [N]/3 | ... |
| Agents | [N]/3 | ... |
| Hooks | [N]/3 | ... |
| Rules | [N]/3 | ... |
| MCP | [N]/3 | ... |
| Eval | [N]/3 | ... |
| Team | [N]/3 | ... |

🔧 Improvement suggestions:
1. [How to improve the lowest-scoring dimension]
2. ...

Context-File Pruning (quarterly routine)

CLAUDE.md is a bias-correction prompt, not a knowledge store — density is quality. Each quarter, classify every line:

Additional checks:

• Patterns corrected 3+ times in tasks/lessons.md → report as promotion candidates for a CLAUDE.md line
• Lines that no longer fire after model generations change (the targeted tendency is gone) → delete candidates
• Output the classification as a table; delete only after user approval