skill-judge_skill

Overview

This skill evaluates agent Skill design quality against official specifications and best practices. It scores multiple dimensions of a Skill package and produces actionable, prioritized improvement suggestions. Use it to audit, review, or harden Skill documentation and packaging for reliable activation and efficient token use.

How this skill works

The skill inspects the Skill package metadata, description, content layering, and documented procedures to measure adherence to official format and practical design patterns. It computes multi-dimensional scores (knowledge delta, mindset/procedures, anti-patterns, description quality, progressive disclosure, freedom calibration) and generates concrete fixes with rationale and severity. Results include a breakdown of weaknesses, exact lines/sections to trim or expand, and suggested trigger keywords and loading guidance.

When to use it

• During pre-release audits of a Skill package
• When improving the activation probability and precision of a Skill
• Prior to bundling Skills into an agent to reduce token waste
• When you need a prioritized remediation plan and example edits
• To train contributors on expert Skill authoring practices

Best practices

• Maximize expert-only content; remove redundant explanations the model already knows
• Write descriptions that state WHAT the Skill does, WHEN to use it, and KEYWORDS to trigger it
• Include decision trees, domain-specific workflows, and explicit anti-patterns with reasons
• Layer content: short metadata, concise body, and on-demand resources with clear loading triggers
• Calibrate specificity to task fragility: constrain fragile operations and allow freedom for creative tasks

Example use cases

• Audit a Skill description to ensure reliable automatic activation and add missing trigger terms
• Refactor a long Skill body into metadata, body, and reference files with mandatory load triggers
• Identify and remove redundant background material to improve context efficiency
• Add expert-only decision trees and NEVER lists to raise the knowledge-delta score
• Produce a prioritized remediation report for maintainers with sample text replacements