AgentSkillsCN

skill-evaluator

使用基于评分标准的评估来评价智能体技能质量。当评估技能有效性、识别改进机会或确保质量标准时使用。

SKILL.md
--- frontmatter
name: skill-evaluator
description: Evaluate agent skill quality using rubric-based assessment. Use when assessing skill effectiveness, identifying improvement opportunities, or ensuring quality standards.
license: MIT
allowed-tools:
  - Read
  - Bash

Skill Evaluator

Assess agent skill quality through comprehensive rubric-based evaluation.

When to Use This Skill

Use skill-evaluator when you need to:

  • •Evaluate a new skill's quality before deployment
  • •Assess an existing skill's effectiveness
  • •Identify specific improvement opportunities
  • •Measure skill quality objectively
  • •Compare skills against quality standards
  • •Generate improvement recommendations

Evaluation Process

Phase 1: Preparation

Gather Context

  1. •Identify the skill to evaluate
  2. •Read the complete SKILL.md
  3. •Review any bundled resources (scripts, templates)
  4. •Understand the skill's intended purpose

Load Evaluation Rubric

bash
cat templates/evaluation-rubric.md

Phase 2: Rubric-Based Evaluation

Evaluate the skill across four dimensions:

1. Clarity (1-5)

What to Assess:

  • •Are instructions imperative and actionable?
  • •Is the description clear about WHAT the skill does?
  • •Is the description clear about WHEN to use it?
  • •Are technical terms explained?
  • •Is the language precise and unambiguous?

Scoring Guide:

  • •5: Exceptionally clear, every instruction actionable, perfect balance of detail
  • •4: Very clear, minor improvements possible
  • •3: Generally clear, some vague sections
  • •2: Unclear in multiple places, hard to follow
  • •1: Confusing, missing critical information

Questions to Ask:

  • •Can someone unfamiliar with the domain understand this?
  • •Are all steps clearly defined?
  • •Is there any ambiguous language?

2. Completeness (1-5)

What to Assess:

  • •Are all necessary steps documented?
  • •Are dependencies clearly stated?
  • •Are edge cases addressed?
  • •Is error handling covered?
  • •Are prerequisites listed?

Scoring Guide:

  • •5: Comprehensive, covers all scenarios, no gaps
  • •4: Very complete, minor edge cases missing
  • •3: Generally complete, some steps assumed
  • •2: Significant gaps, missing key information
  • •1: Incomplete, critical steps missing

Questions to Ask:

  • •Could someone execute this skill without asking questions?
  • •Are dependencies explicit?
  • •What happens when things go wrong?

3. Examples (1-5)

What to Assess:

  • •Are concrete examples provided?
  • •Do examples show expected outcomes?
  • •Are multiple scenarios covered?
  • •Do examples help clarify instructions?
  • •Are examples realistic and practical?

Scoring Guide:

  • •5: Excellent examples, multiple scenarios, clear outcomes
  • •4: Good examples, covers main use cases
  • •3: Basic examples present, could be more detailed
  • •2: Minimal examples, not very helpful
  • •1: No examples or examples are confusing

Questions to Ask:

  • •Do examples demonstrate the skill in action?
  • •Are edge cases illustrated?
  • •Can users adapt examples to their needs?

4. Focus (1-5)

What to Assess:

  • •Does the skill address a specific, well-defined task?
  • •Is it too broad or monolithic?
  • •Does it try to do too much?
  • •Is the scope appropriate?
  • •Are responsibilities clear?

Scoring Guide:

  • •5: Perfectly focused, single clear purpose
  • •4: Well-focused, minor scope creep
  • •3: Generally focused, some unnecessary breadth
  • •2: Too broad, trying to do too much
  • •1: Unfocused, unclear purpose

Questions to Ask:

  • •Can you describe the skill's purpose in one sentence?
  • •Should this be split into multiple skills?
  • •Is anything out of scope included?

Phase 3: Scoring

Calculate Scores

For each dimension:

  1. •Review the assessment criteria
  2. •Assign a score (1-5)
  3. •Document specific observations
  4. •Note evidence supporting the score

Total Score

Sum all dimensions (max: 20)

Quality Thresholds

  • •18-20: Excellent - Production ready
  • •15-17: Strong - Minor improvements recommended
  • •12-14: Adequate - Needs refinement
  • •9-11: Weak - Significant improvements required
  • •1-8: Poor - Major rework needed

Phase 4: Recommendations

Generate Actionable Feedback

For each dimension with score <5:

  1. •Identify specific weaknesses
  2. •Provide concrete improvement suggestions
  3. •Prioritize recommendations (high/medium/low)

Example Recommendations:

Clarity (3/5):

  • •High: Rewrite Phase 2 instructions in imperative form
  • •Medium: Define technical term "API endpoint" in context
  • •Low: Add section headings for better scanning

Completeness (4/5):

  • •Medium: Add error handling section
  • •Low: List Python version requirement

Examples (3/5):

  • •High: Add 2 more concrete examples showing different scenarios
  • •Medium: Include expected output for Example 1
  • •Medium: Add edge case example (empty input)

Focus (5/5):

  • •No recommendations - well-focused

Phase 5: Generate Report

Create Evaluation Report

Use the rubric template:

bash
cat templates/evaluation-rubric.md

Fill in:

  • •Skill name and date
  • •Scores for each dimension
  • •Total score and quality level
  • •Detailed observations
  • •Prioritized recommendations
  • •Overall assessment

Report Structure

markdown
# Skill Evaluation Report

**Skill**: skill-name
**Date**: YYYY-MM-DD
**Evaluator**: [Agent/Human name]

## Scores

| Dimension    | Score     | Notes                          |
| ------------ | --------- | ------------------------------ |
| Clarity      | 4/5       | Very clear, minor improvements |
| Completeness | 5/5       | Comprehensive coverage         |
| Examples     | 3/5       | Needs more concrete examples   |
| Focus        | 5/5       | Well-scoped and focused        |
| **Total**    | **17/20** | **Strong**                     |

## Detailed Assessment

[Per-dimension analysis]

## Recommendations

[Prioritized improvement suggestions]

## Overall Assessment

[Summary and final thoughts]

Phase 6: Iterate

Support Improvement Cycle

  1. •Share evaluation report with skill author
  2. •Implement high-priority recommendations
  3. •Re-evaluate after changes
  4. •Compare scores to measure improvement
  5. •Iterate until quality threshold met

Evaluation Rubric

Use the bundled rubric template:

bash
cat templates/evaluation-rubric.md

The template provides:

  • •Detailed scoring criteria for each dimension
  • •Example assessments
  • •Report structure
  • •Question prompts

Scripts

run_evaluation.py

Purpose: Semi-automated evaluation assistance

Usage:

bash
python scripts/run_evaluation.py /path/to/skill-directory

Features:

  • •Extracts skill metadata
  • •Calculates word counts
  • •Identifies missing sections
  • •Generates report template
  • •Assists with objective metrics

Note: Final scoring requires human/agent judgment

Examples

Example 1: High-Quality Skill Evaluation

Skill: brand-guidelines

Assessment:

Clarity: 5/5

  • •Description perfectly explains what and when
  • •Instructions are imperative and actionable
  • •Technical terms clear from context
  • •Language is precise

Completeness: 5/5

  • •All necessary information present
  • •Color codes, typography, spacing all specified
  • •No external dependencies needed
  • •Self-contained and complete

Examples: 4/5

  • •Color examples provided with hex codes
  • •Typography examples with sizes and weights
  • •Could benefit from visual mockup examples

Focus: 5/5

  • •Single clear purpose: brand guidelines
  • •Not trying to do design work, just provide standards
  • •Perfect scope

Total: 19/20 (Excellent)

Recommendations:

  • •Low: Add 1-2 visual mockup examples showing guidelines applied

Example 2: Skill Needing Improvement

Skill: code-helper

Assessment:

Clarity: 2/5

  • •Description vague: "Helps with code"
  • •Instructions use passive voice
  • •Many undefined terms
  • •Ambiguous steps

Completeness: 3/5

  • •Missing dependency information
  • •No error handling guidance
  • •Prerequisites not stated
  • •Some steps documented

Examples: 2/5

  • •Only one example provided
  • •Example doesn't show expected output
  • •No edge cases illustrated

Focus: 2/5

  • •Tries to do too much: formatting, analysis, documentation
  • •Should be 3 separate skills
  • •Unclear primary purpose

Total: 9/20 (Weak)

Recommendations:

High Priority:

  • •Split into 3 focused skills: code-formatter, code-analyzer, code-documenter
  • •Rewrite description to explain specific purpose and when to use
  • •Rewrite all instructions in imperative form

Medium Priority:

  • •Add 3-4 concrete examples with inputs and outputs
  • •Document all dependencies (Python version, packages)
  • •Add error handling section

Low Priority:

  • •Define all technical terms in context
  • •Add edge case examples

Example 3: Re-Evaluation After Improvements

Initial Evaluation: 12/20 (Adequate) After Improvements: 17/20 (Strong)

Changes Made:

  • •Added 3 concrete examples (+2 points)
  • •Clarified description to explain when to use (+1 point)
  • •Rewrote vague instructions in imperative form (+2 points)

Impact: Moved from "Needs refinement" to "Minor improvements recommended"

Best Practices

Be Objective

  • •Use the rubric consistently
  • •Base scores on evidence, not feelings
  • •Compare against criteria, not other skills

Be Specific

  • •Document exact issues
  • •Provide concrete examples of problems
  • •Suggest specific improvements

Be Constructive

  • •Focus on helping improve the skill
  • •Acknowledge what works well
  • •Prioritize feedback (don't overwhelm)

Be Thorough

  • •Read the entire skill before scoring
  • •Check all bundled resources
  • •Consider real-world usage

Common Evaluation Pitfalls

Leniency Bias

❌ Scoring too high because skill "seems okay" ✓ Compare against explicit rubric criteria

Recency Effect

❌ Letting recent sections heavily influence score ✓ Evaluate each dimension independently

Halo Effect

❌ One strong area influences scoring of others ✓ Score each dimension separately

Comparison Bias

❌ Scoring relative to other skills ✓ Score against absolute rubric criteria

Quality Gates

Minimum Thresholds:

  • •For deployment: ≥15/20 total, no dimension <3
  • •For production: ≥18/20 total, all dimensions ≥4
  • •Best practice: All dimensions ≥4

Red Flags (immediate revision needed):

  • •Any dimension scored 1
  • •Total score <12
  • •Missing examples entirely
  • •Vague description

Evaluation Checklist

Before finalizing evaluation:

  • • Read complete SKILL.md
  • • Reviewed all bundled resources
  • • Scored all four dimensions
  • • Calculated total score
  • • Documented specific observations
  • • Generated prioritized recommendations
  • • Identified quality threshold met
  • • Created evaluation report

Templates

Evaluation Rubric Template

bash
cat templates/evaluation-rubric.md

Provides:

  • •Scoring criteria details
  • •Report structure
  • •Example assessments

Resources