Skills eval

Skills Evaluation and Improvement

When NOT To Use

• Writing a new skill (use abstract:skill-authoring)
• Evaluating hooks (use abstract:hooks-eval)
• Evaluating rules (use abstract:rules-eval)

Table of Contents

1. Overview
2. Quick Start
3. Evaluation Workflow
4. Evaluation and Optimization
5. Resources

Overview

This framework audits Claude skills against quality standards to improve performance and reduce token consumption. Automated tools analyze skill structure, measure context usage, and identify specific technical improvements. Run verification commands after each audit to confirm fixes work correctly.

The skills-auditor provides structural analysis, while the improvement-suggester ranks fixes by impact. Compliance is verified through the compliance-checker. Runtime efficiency is monitored by tool-performance-analyzer and token-usage-tracker.

Quick Start

Basic Audit

Run a full audit of all skills or target a specific file to identify structural issues.

# Audit all skills
make audit-all

# Audit specific skill
make audit-skill TARGET=path/to/skill/SKILL.md

Analysis and Optimization

Use skill_analyzer.py for complexity checks and token_estimator.py to verify the context budget.

make analyze-skill TARGET=path/to/skill/SKILL.md
make estimate-tokens TARGET=path/to/skill/SKILL.md

Improvements

Generate a prioritized plan and verify standards compliance using improvement_suggester.py and compliance_checker.py.

make improve-skill TARGET=path/to/skill/SKILL.md
make check-compliance TARGET=path/to/skill/SKILL.md

Evaluation Workflow

Start with make audit-all to inventory skills and identify high-priority targets. For each skill requiring attention, run analysis with analyze-skill to map complexity. Generate an improvement plan, apply fixes, and run check-compliance to verify the skill meets project standards. Finalize by checking the token budget for efficiency.

Evaluation and Optimization

Quality assessments use the skills-auditor and improvement-suggester to generate detailed reports. Performance analysis focuses on token efficiency through the token-usage-tracker and tool performance via tool-performance-analyzer. For standards compliance, the compliance-checker automates common fixes for structural issues.

Scoring and Prioritization

We evaluate skills across five dimensions: structure compliance, content quality, token efficiency, activation reliability, and tool integration. Scores above 90 represent production-ready skills, while scores below 50 indicate critical issues requiring immediate attention.

Improvements are prioritized by impact. Critical issues include security vulnerabilities or broken functionality. High-priority items cover structural flaws that hinder discoverability. Medium and low priorities focus on best practices and minor optimizations.

Structural Patterns

Deprecated: skills/shared/modules/ directories. Shared modules must be relocated into the consuming skill's own modules/ directory. The evaluator flags any remaining skills/shared/ as a structural warning.

Current: Each skill owns its modules at skills/<skill-name>/modules/. Cross-skill references use relative paths (e.g., ../skill-authoring/modules/anti-rationalization.md).

Resources

Shared Modules: Cross-Skill Patterns

• Anti-Rationalization Patterns: See anti-rationalization.md
• Enforcement Language: See enforcement-language.md
• Trigger Patterns: See trigger-patterns.md

Skill-Specific Modules

• Trigger Isolation Analysis: See modules/trigger-isolation-analysis.md
• Authoring Checklist: See modules/authoring-checklist.md
• Evaluation Workflows: See modules/evaluation-workflows.md
• Advanced Tool Use Analysis: See modules/advanced-tool-use-analysis.md
• Evaluation Framework: See modules/evaluation-framework.md
• Integration Patterns: See modules/integration.md
• Troubleshooting: See modules/troubleshooting.md
• Pressure Testing: See modules/pressure-testing.md
• Integration Testing: See modules/integration-testing.md
• Performance Benchmarking: See modules/performance-benchmarking.md

Tools and Automation

• Tools: Executable analysis utilities in scripts/ directory.
• Automation: Setup and validation scripts in scripts/automation/.

Exit Criteria

• Every audited skill receives a score across all five dimensions (structure compliance, content quality, token efficiency, activation reliability, tool integration) summing to 100.
• Any skill with a deprecated skills/shared/ module reference is listed as a structural warning in the audit output.
• make check-compliance TARGET=<skill> exits 0 for skills reported as passing, confirming the compliance-checker agrees with the audit score.
• An improvement plan is produced for any skill scoring below 75, with findings ordered by priority: critical > high > medium > low.