Skill Tuning

Skill Tuning

Autonomous diagnosis and optimization for skill execution issues.

Pre-load (before execution)

1. Codebase docs: If .workflow/codebase/ARCHITECTURE.md exists, read for project context
2. Specs: maestro load --type spec --category coding — load coding conventions
3. Wiki knowledge: maestro search "skill design optimization" --json — top 5 entries as prior context
4. All optional — proceed without if unavailable

Architecture

┌─────────────────────────────────────────────────────┐
│  Phase 0: Read Specs (mandatory)                    │
│  → problem-taxonomy.md, tuning-strategies.md         │
└─────────────────────────────────────────────────────┘
                        ↓
┌─────────────────────────────────────────────────────┐
│  Orchestrator (state-driven)                         │
│  Read state → Select action → Execute → Update → ✓ │
└─────────────────────────────────────────────────────┘
        ↓                           ↓
┌──────────────────────┐   ┌──────────────────┐
│  Diagnosis Phase     │   │ Agy CLI       │
│  • Context          │   │ Deep analysis    │
│  • Memory           │   │ (on-demand)      │
│  • DataFlow         │   │                  │
│  • Agent            │   │ Complex issues   │
│  • Docs             │   │ Architecture     │
│  • Token Usage      │   │ Performance      │
└──────────────────────┘   └──────────────────┘
                ↓
        ┌───────────────────┐
        │  Fix & Verify     │
        │  Apply → Re-test  │
        └───────────────────┘

Core Issues Detected

Problem Categories (Detailed Specs)

See specs/problem-taxonomy.md for:

• Detection patterns (regex/checks)
• Severity calculations
• Impact assessments

Tuning Strategies (Detailed Specs)

See specs/tuning-strategies.md for:

• 10+ strategies per category
• Implementation patterns
• Verification methods

Workflow

Action Reference

Full action details: phases/actions/

State Management

Single source of truth: {run_dir}/outputs/skill-tuning-{ts}/state.json

{
  "status": "pending|running|completed|failed",
  "target_skill": { "name": "...", "path": "..." },
  "diagnosis": {
    "context": {...},
    "memory": {...},
    "dataflow": {...},
    "agent": {...},
    "docs": {...},
    "token_consumption": {...}
  },
  "issues": [{"id":"...", "severity":"...", "category":"...", "strategy":"..."}],
  "proposed_fixes": [...],
  "applied_fixes": [...],
  "quality_gate": "pass|fail",
  "final_report": "..."
}

See phases/state-schema.md for complete schema.

Orchestrator Logic

See phases/orchestrator.md for:

• Decision logic (termination checks → action selection)
• State transitions
• Error recovery

Key Principles

1. Problem-First: Diagnosis before any fix
2. Data-Driven: Record traces, token counts, snapshots
3. Iterative: Multiple rounds until quality gates pass
4. Reversible: All changes with backup checkpoints
5. Non-Invasive: Minimal changes, maximum clarity

Usage Examples

# Basic skill diagnosis
/skill-tuning "Fix memory leaks in my skill"

# Deep analysis with Agy
/skill-tuning "Architecture issues in async workflow"

# Focus on specific areas
/skill-tuning "Optimize token consumption and fix agent coordination"

# Custom issue
/skill-tuning "My skill produces inconsistent outputs"

Output

After completion, review:

• {run_dir}/outputs/skill-tuning-{ts}/state.json - Full state with final_report
• state.final_report - Markdown summary (in state.json)
• state.applied_fixes - List of applied fixes with verification results

Reference Documents