skill-judge

Skill Judge

Evaluate Agent Skills against official specifications and patterns derived from 17+ official examples.

---

Core Philosophy

What is a Skill?

A Skill is NOT a tutorial. A Skill is a knowledge externalization mechanism.

Traditional AI knowledge is locked in model parameters. To teach new capabilities:

Traditional: Collect data → GPU cluster → Train → Deploy new version
Cost: $10,000 - $1,000,000+
Timeline: Weeks to months

Skills change this:

Skill: Edit SKILL.md → Save → Takes effect on next invocation
Cost: $0
Timeline: Instant

This is the paradigm shift from "training AI" to "educating AI" — like a hot-swappable LoRA adapter that requires no training. You edit a Markdown file in natural language, and the model's behavior changes.

The Core Formula

Good Skill = Expert-only Knowledge − What Claude Already Knows

A Skill's value is measured by its knowledge delta — the gap between what it provides and what the model already knows.

• Expert-only knowledge: Decision trees, trade-offs, edge cases, anti-patterns, domain-specific thinking frameworks — things that take years of experience to accumulate
• What Claude already knows: Basic concepts, standard library usage, common programming patterns, general best practices

When a Skill explains "what is PDF" or "how to write a for-loop", it's compressing knowledge Claude already has. This is token waste — context window is a public resource shared with system prompts, conversation history, other Skills, and user requests.

Tool vs Skill

Concept	Essence	Function	Example
Tool	What model CAN do	Execute actions	bash, read_file, write_file, WebSearch
Skill	What model KNOWS how to do	Guide decisions	PDF processing, MCP building, frontend design

Tools define capability boundaries — without bash tool, model can't execute commands. Skills inject knowledge — without frontend-design Skill, model produces generic UI.

The equation:

General Agent + Excellent Skill = Domain Expert Agent

Same Claude model, different Skills loaded, becomes different experts.

Three Types of Knowledge in Skills

When evaluating, categorize each section:

Type	Definition	Treatment
Expert	Claude genuinely doesn't know this	Must keep — this is the Skill's value
Activation	Claude knows but may not think of	Keep if brief — serves as reminder
Redundant	Claude definitely knows this	Should delete — wastes tokens

The art of Skill design is maximizing Expert content, using Activation sparingly, and eliminating Redundant ruthlessly.

---

Evaluation Dimensions (120 points total)

D1: Knowledge Delta (20 points) — THE CORE DIMENSION

The most important dimension. Does the Skill add genuine expert knowledge?

Score	Criteria
0-5	Explains basics Claude knows (what is X, how to write code, standard library tutorials)
6-10	Mixed: some expert knowledge diluted by obvious content
11-15	Mostly expert knowledge with minimal redundancy
16-20	Pure knowledge delta — every paragraph earns its tokens

Red flags (instant score ≤5):

• "What is [basic concept]" sections
• Step-by-step tutorials for standard operations
• Explaining how to use common libraries
• Generic best practices ("write clean code", "handle errors")
• Definitions of industry-standard terms

Green flags (indicators of high knowledge delta):

• Decision trees for non-obvious choices ("when X fails, try Y because Z")
• Trade-offs only an expert would know ("A is faster but B handles edge case C")
• Edge cases from real-world experience
• "NEVER do X because [non-obvious reason]"
• Domain-specific thinking frameworks

Evaluation questions:

1. For each section, ask: "Does Claude already know this?"
2. If explaining something, ask: "Is this explaining TO Claude or FOR Claude?"
3. Count paragraphs that are Expert vs Activation vs Redundant

---

D2: Mindset + Appropriate Procedures (15 points)

Does the Skill transfer expert thinking patterns along with necessary domain-specific procedures?

The difference between experts and novices isn't "knowing how to operate" — it's "how to think about the problem." But thinking patterns alone aren't enough when Claude lacks domain-specific procedural knowledge.

Key distinction:

Type	Example	Value
Thinking patterns	"Before designing, ask: What makes this memorable?"	High — shapes decision-making
Domain-specific procedures	"OOXML workflow: unpack → edit XML → validate → pack"	High — Claude may not know this
Generic procedures	"Step 1: Open file, Step 2: Edit, Step 3: Save"	Low — Claude already knows

Score	Criteria
0-3	Only generic procedures Claude already knows
4-7	Has domain procedures but lacks thinking frameworks
8-11	Good balance: thinking patterns + domain-specific workflows
12-15	Expert-level: shapes thinking AND provides procedures Claude wouldn't know

What counts as valuable procedures:

• Workflows Claude hasn't been trained on (new tools, proprietary systems)
• Correct ordering that's non-obvious (e.g., "validate BEFORE packing, not after")
• Critical steps that are easy to miss (e.g., "MUST recalculate formulas after editing")
• Domain-specific sequences (e.g., MCP server's 4-phase development process)

What counts as redundant procedures:

• Generic file operations (open, read, write, save)
• Standard programming patterns (loops, conditionals, error handling)
• Common library usage that's well-documented

Expert thinking patterns look like:

Before [action], ask yourself:
- **Purpose**: What problem does this solve? Who uses it?
- **Constraints**: What are the hidden requirements?
- **Differentiation**: What makes this solution memorable?

Valuable domain procedures look like:

### Redlining Workflow (Claude wouldn't know this sequence)
1. Convert to markdown: `pandoc --track-changes=all`
2. Map text to XML: grep for text in document.xml
3. Implement changes in batches of 3-10
4. Pack and verify: check ALL changes were applied

Redundant generic procedures look like:

Step 1: Open the file
Step 2: Find the section
Step 3: Make the change
Step 4: Save and test

The test:

1. Does it tell Claude WHAT to think about? (thinking patterns)
2. Does it tell Claude HOW to do things it wouldn't know? (domain procedures)

A good Skill provides both when needed.

---

D3: Anti-Pattern Quality (15 points)

Does the Skill have effective NEVER lists?

Why this matters: Half of expert knowledge is knowing what NOT to do. A senior designer sees purple gradient on white background and instinctively cringes — "too AI-generated." This intuition for "what absolutely not to do" comes from stepping on countless landmines.

Claude hasn't stepped on these landmines. It doesn't know Inter font is overused, doesn't know purple gradients are the signature of AI-generated content. Good Skills must explicitly state these "absolute don'ts."

Score	Criteria
0-3	No anti-patterns mentioned
4-7	Generic warnings ("avoid errors", "be careful", "consider edge cases")
8-11	Specific NEVER list with some reasoning
12-15	Expert-grade anti-patterns with WHY — things only experience teaches

Expert anti-patterns (specific + reason):

NEVER use generic AI-generated aesthetics like:
- Overused font families (Inter, Roboto, Arial)
- Cliched color schemes (particularly purple gradients on white backgrounds)
- Predictable layouts and component patterns
- Default border-radius on everything

Weak anti-patterns (vague, no reasoning):

Avoid making mistakes.
Be careful with edge cases.
Don't write bad code.

The test: Would an expert read the anti-pattern list and say "yes, I learned this the hard way"? Or would they say "this is obvious to everyone"?

---

D4: Specification Compliance — Especially Description (15 points)

Does the Skill follow official format requirements? Special focus on description quality.

Score	Criteria
0-5	Missing frontmatter or invalid format
6-10	Has frontmatter but description is vague or incomplete
11-13	Valid frontmatter, description has WHAT but weak on WHEN
14-15	Perfect: comprehensive description with WHAT, WHEN, and trigger keywords

Frontmatter requirements:

• name: lowercase, alphanumeric + hyphens only, ≤64 characters
• description: THE MOST CRITICAL FIELD — determines if skill gets used at all

---

Why description is THE MOST IMPORTANT field:

┌─────────────────────────────────────────────────────────────────────┐
│  SKILL ACTIVATION FLOW                                              │
│                                                                     │
│  User Request → Agent sees ALL skill descriptions → Decides which  │
│                 (only descriptions, not bodies!)     to activate    │
│                                                                     │
│  If description doesn't match → Skill NEVER gets loaded            │
│  If description is vague → Skill might not trigger when it should  │
│  If description lacks keywords → Skill is invisible to the Agent   │
└─────────────────────────────────────────────────────────────────────┘

The brutal truth: A Skill with perfect content but poor description is useless — it will never be activated. The description is the only chance to tell the Agent "use me in these situations."

---

Description must answer THREE questions:

1. WHAT: What does this Skill do? (functionality)
2. WHEN: In what situations should it be used? (trigger scenarios)
3. KEYWORDS: What terms should trigger this Skill? (searchable terms)

Excellent description (all three elements):

description: "Comprehensive document creation, editing, and analysis with support
for tracked changes, comments, formatting preservation, and text extraction.
When Claude needs to work with professional documents (.docx files) for:
(1) Creating new documents, (2) Modifying or editing content,
(3) Working with tracked changes, (4) Adding comments, or any other document tasks"

Analysis:

• WHAT: creation, editing, analysis, tracked changes, comments
• WHEN: "When Claude needs to work with... for: (1)... (2)... (3)..."
• KEYWORDS: .docx files, tracked changes, professional documents

Poor description (missing elements):

description: "处理文档相关功能"

Problems:

• WHAT: vague ("文档相关功能" — what specifically?)
• WHEN: missing (when should Agent use this?)
• KEYWORDS: missing (no ".docx", no specific scenarios)

Another poor example:

description: "A helpful skill for various tasks"

This is useless — Agent has no idea when to activate it.

---

Description quality checklist:

• Lists specific capabilities (not just "helps with X")
• Includes explicit trigger scenarios ("Use when...", "When user asks for...")
• Contains searchable keywords (file extensions, domain terms, action verbs)
• Specific enough that Agent knows EXACTLY when to use it
• Includes scenarios where this skill MUST be used (not just "can be used")

---

D5: Progressive Disclosure (15 points)

Does the Skill implement proper content layering?

Skill loading has three layers:

Layer 1: Metadata (always in memory)
         Only name + description
         ~100 tokens per skill

Layer 2: SKILL.md Body (loaded after triggering)
         Detailed guidelines, code examples, decision trees
         Ideal: < 500 lines

Layer 3: Resources (loaded on demand)
         scripts/, references/, assets/
         No limit

Score	Criteria
0-5	Everything dumped in SKILL.md (>500 lines, no structure)
6-10	Has references but unclear when to load them
11-13	Good layering with MANDATORY triggers present
14-15	Perfect: decision trees + explicit triggers + "Do NOT Load" guidance

For Skills WITH references directory, check Loading Trigger Quality:

Trigger Quality	Characteristics
Poor	References listed at end, no loading guidance
Mediocre	Some triggers but not embedded in workflow
Good	MANDATORY triggers in workflow steps
Excellent	Scenario detection + conditional triggers + "Do NOT Load"

The loading problem:

Loading too little ◄─────────────────────────────────► Loading too much
- References sit unused                    - Wastes context space
- Agent doesn't know when to load          - Irrelevant info dilutes key content
- Knowledge is there but never accessed    - Unnecessary token overhead

Good loading trigger (embedded in workflow):

### Creating New Document

**MANDATORY - READ ENTIRE FILE**: Before proceeding, you MUST read
[`docx-js.md`](docx-js.md) (~500 lines) completely from start to finish.
**NEVER set any range limits when reading this file.**

**Do NOT load** `ooxml.md` or `redlining.md` for this task.

Bad loading trigger (just listed):

## References
- docx-js.md - for creating documents
- ooxml.md - for editing
- redlining.md - for tracking changes

For simple Skills (no references, <100 lines): Score based on conciseness and self-containment.

---

D6: Freedom Calibration (15 points)

Is the level of specificity appropriate for the task's fragility?

Different tasks need different levels of constraint. This is about matching freedom to fragility.

Score	Criteria
0-5	Severely mismatched (rigid scripts for creative tasks, vague for fragile ops)
6-10	Partially appropriate, some mismatches
11-13	Good calibration for most scenarios
14-15	Perfect freedom calibration throughout

The freedom spectrum:

Task Type	Should Have	Why	Example Skill
Creative/Design	High freedom	Multiple valid approaches, differentiation is value	frontend-design
Code review	Medium freedom	Principles exist but judgment required	code-review
File format operations	Low freedom	One wrong byte corrupts file, consistency critical	docx, xlsx, pdf

High freedom (text-based instructions):

Commit to a BOLD aesthetic direction. Pick an extreme: brutally minimal,
maximalist chaos, retro-futuristic, organic natural...

Medium freedom (pseudocode or parameterized):

Review priority:
1. Security vulnerabilities (must fix)
2. Logic errors (must fix)
3. Performance issues (should fix)
4. Maintainability (optional)

Low freedom (specific scripts, exact steps):

**MANDATORY**: Use exact script in `scripts/create-doc.py`
Parameters: --title "X" --author "Y"
Do NOT modify the script.

The test: Ask "if Agent makes a mistake, what's the consequence?"

• High consequence → Low freedom
• Low consequence → High freedom

---

D7: Pattern Recognition (10 points)

Does the Skill follow an established official pattern?

Through analyzing 17 official Skills, we identified 5 main design patterns:

Pattern	~Lines	Key Characteristics	Example	When to Use
Mindset	~50	Thinking > technique, strong NEVER list, high freedom	frontend-design	Creative tasks requiring taste
Navigation	~30	Minimal SKILL.md, routes to sub-files	internal-comms	Multiple distinct scenarios
Philosophy	~150	Two-step: Philosophy → Express, emphasizes craft	canvas-design	Art/creation requiring originality
Process	~200	Phased workflow, checkpoints, medium freedom	mcp-builder	Complex multi-step projects
Tool	~300	Decision trees, code examples, low freedom	docx, pdf, xlsx	Precise operations on specific formats

Score	Criteria
0-3	No recognizable pattern, chaotic structure
4-6	Partially follows a pattern with significant deviations
7-8	Clear pattern with minor deviations
9-10	Masterful application of appropriate pattern

Pattern selection guide:

Your Task Characteristics	Recommended Pattern
Needs taste and creativity	Mindset (~50 lines)
Needs originality and craft quality	Philosophy (~150 lines)
Has multiple distinct sub-scenarios	Navigation (~30 lines)
Complex multi-step project	Process (~200 lines)
Precise operations on specific format	Tool (~300 lines)

---

D8: Practical Usability (15 points)

Can an Agent actually use this Skill effectively?

Score	Criteria
0-5	Confusing, incomplete, contradictory, or untested guidance
6-10	Usable but with noticeable gaps
11-13	Clear guidance for common cases
14-15	Comprehensive coverage including edge cases and error handling

Check for:

• Decision trees: For multi-path scenarios, is there clear guidance on which path to take?
• Code examples: Do they actually work? Or are they pseudocode that breaks?
• Error handling: What if the main approach fails? Are fallbacks provided?
• Edge cases: Are unusual but realistic scenarios covered?
• Actionability: Can Agent immediately act, or needs to figure things out?

Good usability (decision tree + fallback):

| Task | Primary Tool | Fallback | When to Use Fallback |
|------|-------------|----------|----------------------|
| Read text | pdftotext | PyMuPDF | Need layout info |
| Extract tables | camelot-py | tabula-py | camelot fails |

**Common issues**:
- Scanned PDF: pdftotext returns blank → Use OCR first
- Encrypted PDF: Permission error → Use PyMuPDF with password

Poor usability (vague):

Use appropriate tools for PDF processing.
Handle errors properly.
Consider edge cases.

---

NEVER Do When Evaluating

• NEVER give high scores just because it "looks professional" or is well-formatted
• NEVER ignore token waste — every redundant paragraph should result in deduction
• NEVER let length impress you — a 43-line Skill can outperform a 500-line Skill
• NEVER skip mentally testing the decision trees — do they actually lead to correct choices?
• NEVER forgive explaining basics with "but it provides helpful context"
• NEVER overlook missing anti-patterns — if there's no NEVER list, that's a significant gap
• NEVER assume all procedures are valuable — distinguish domain-specific from generic
• NEVER undervalue the description field — poor description = skill never gets used
• NEVER put "when to use" info only in the body — Agent only sees description before loading

---

Evaluation Protocol

Step 1: First Pass — Knowledge Delta Scan

Read SKILL.md completely and for each section ask:

"Does Claude already know this?"

Mark each section as:

• [E] Expert: Claude genuinely doesn't know this — value-add
• [A] Activation: Claude knows but brief reminder is useful — acceptable
• [R] Redundant: Claude definitely knows this — should be deleted

Calculate rough ratio: E:A:R

• Good Skill: >70% Expert, <20% Activation, <10% Redundant
• Mediocre Skill: 40-70% Expert, high Activation
• Bad Skill: <40% Expert, high Redundant

Step 2: Structure Analysis

[ ] Check frontmatter validity
[ ] Count total lines in SKILL.md
[ ] List all reference files and their sizes
[ ] Identify which pattern the Skill follows
[ ] Check for loading triggers (if references exist)

Step 3: Score Each Dimension

For each of the 8 dimensions:

1. Find specific evidence (quote relevant lines)
2. Assign score with one-line justification
3. Note specific improvements if score < max

Step 4: Calculate Total & Grade

Total = D1 + D2 + D3 + D4 + D5 + D6 + D7 + D8
Max = 120 points

Grade Scale (percentage-based):

Grade	Percentage	Meaning
A	90%+ (108+)	Excellent — production-ready expert Skill
B	80-89% (96-107)	Good — minor improvements needed
C	70-79% (84-95)	Adequate — clear improvement path
D	60-69% (72-83)	Below Average — significant issues
F	<60% (<72)	Poor — needs fundamental redesign

Step 5: Generate Report

# Skill Evaluation Report: [Skill Name]

## Summary
- **Total Score**: X/120 (X%)
- **Grade**: [A/B/C/D/F]
- **Pattern**: [Mindset/Navigation/Philosophy/Process/Tool]
- **Knowledge Ratio**: E:A:R = X:Y:Z
- **Verdict**: [One sentence assessment]

## Dimension Scores

| Dimension | Score | Max | Notes |
|-----------|-------|-----|-------|
| D1: Knowledge Delta | X | 20 | |
| D2: Mindset vs Mechanics | X | 15 | |
| D3: Anti-Pattern Quality | X | 15 | |
| D4: Specification Compliance | X | 15 | |
| D5: Progressive Disclosure | X | 15 | |
| D6: Freedom Calibration | X | 15 | |
| D7: Pattern Recognition | X | 10 | |
| D8: Practical Usability | X | 15 | |

## Critical Issues
[List must-fix problems that significantly impact the Skill's effectiveness]

## Top 3 Improvements
1. [Highest impact improvement with specific guidance]
2. [Second priority improvement]
3. [Third priority improvement]

## Detailed Analysis
[For each dimension scoring below 80%, provide:
- What's missing or problematic
- Specific examples from the Skill
- Concrete suggestions for improvement]

---

Common Failure Patterns

Pattern 1: The Tutorial

Symptom: Explains what PDF is, how Python works, basic library usage
Root cause: Author assumes Skill should "teach" the model
Fix: Claude already knows this. Delete all basic explanations.
     Focus on expert decisions, trade-offs, and anti-patterns.

Pattern 2: The Dump

Symptom: SKILL.md is 800+ lines with everything included
Root cause: No progressive disclosure design
Fix: Core routing and decision trees in SKILL.md (<300 lines ideal)
     Detailed content in references/, loaded on-demand

Pattern 3: The Orphan References

Symptom: References directory exists but files are never loaded
Root cause: No explicit loading triggers
Fix: Add "MANDATORY - READ ENTIRE FILE" at workflow decision points
     Add "Do NOT Load" to prevent over-loading

Pattern 4: The Checkbox Procedure

Symptom: Step 1, Step 2, Step 3... mechanical procedures
Root cause: Author thinks in procedures, not thinking frameworks
Fix: Transform into "Before doing X, ask yourself..."
     Focus on decision principles, not operation sequences

Pattern 5: The Vague Warning

Symptom: "Be careful", "avoid errors", "consider edge cases"
Root cause: Author knows things can go wrong but hasn't articulated specifics
Fix: Specific NEVER list with concrete examples and non-obvious reasons
     "NEVER use X because [specific problem that takes experience to learn]"

Pattern 6: The Invisible Skill

Symptom: Great content but skill rarely gets activated
Root cause: Description is vague, missing keywords, or lacks trigger scenarios
Fix: Description must answer WHAT, WHEN, and include KEYWORDS
     "Use when..." + specific scenarios + searchable terms

Example fix:
BAD:  "Helps with document tasks"
GOOD: "Create, edit, and analyze .docx files. Use when working with
       Word documents, tracked changes, or professional document formatting."

Pattern 7: The Wrong Location

Symptom: "When to use this Skill" section in body, not in description
Root cause: Misunderstanding of three-layer loading
Fix: Move all triggering information to description field
     Body is only loaded AFTER triggering decision is made

Pattern 8: The Over-Engineered

Symptom: README.md, CHANGELOG.md, INSTALLATION_GUIDE.md, CONTRIBUTING.md
Root cause: Treating Skill like a software project
Fix: Delete all auxiliary files. Only include what Agent needs for the task.
     No documentation about the Skill itself.

Pattern 9: The Freedom Mismatch

Symptom: Rigid scripts for creative tasks, vague guidance for fragile operations
Root cause: Not considering task fragility
Fix: High freedom for creative (principles, not steps)
     Low freedom for fragile (exact scripts, no parameters)

---

Quick Reference Checklist

┌─────────────────────────────────────────────────────────────────────────┐
│  SKILL EVALUATION QUICK CHECK                                           │
├─────────────────────────────────────────────────────────────────────────┤
│                                                                         │
│  KNOWLEDGE DELTA (most important):                                      │
│    [ ] No "What is X" explanations for basic concepts                   │
│    [ ] No step-by-step tutorials for standard operations                │
│    [ ] Has decision trees for non-obvious choices                       │
│    [ ] Has trade-offs only experts would know                           │
│    [ ] Has edge cases from real-world experience                        │
│                                                                         │
│  MINDSET + PROCEDURES:                                                  │
│    [ ] Transfers thinking patterns (how to think about problems)        │
│    [ ] Has "Before doing X, ask yourself..." frameworks                 │
│    [ ] Includes domain-specific procedures Claude wouldn't know         │
│    [ ] Distinguishes valuable procedures from generic ones              │
│                                                                         │
│  ANTI-PATTERNS:                                                         │
│    [ ] Has explicit NEVER list                                          │
│    [ ] Anti-patterns are specific, not vague                            │
│    [ ] Includes WHY (non-obvious reasons)                               │
│                                                                         │
│  SPECIFICATION (description is critical!):                              │
│    [ ] Valid YAML frontmatter                                           │
│    [ ] name: lowercase, ≤64 chars                                       │
│    [ ] description answers: WHAT does it do?                            │
│    [ ] description answers: WHEN should it be used?                     │
│    [ ] description contains trigger KEYWORDS                            │
│    [ ] description is specific enough for Agent to know when to use     │
│                                                                         │
│  STRUCTURE:                                                             │
│    [ ] SKILL.md < 500 lines (ideal < 300)                               │
│    [ ] Heavy content in references/                                     │
│    [ ] Loading triggers embedded in workflow                            │
│    [ ] Has "Do NOT Load" for preventing over-loading                    │
│                                                                         │
│  FREEDOM:                                                               │
│    [ ] Creative tasks → High freedom (principles)                       │
│    [ ] Fragile operations → Low freedom (exact scripts)                 │
│                                                                         │
│  USABILITY:                                                             │
│    [ ] Decision trees for multi-path scenarios                          │
│    [ ] Working code examples                                            │
│    [ ] Error handling and fallbacks                                     │
│    [ ] Edge cases covered                                               │
│                                                                         │
└─────────────────────────────────────────────────────────────────────────┘

---

The Meta-Question

When evaluating any Skill, always return to this fundamental question:

"Would an expert in this domain, looking at this Skill, say: 'Yes, this captures knowledge that took me years to learn'?"

If the answer is yes → the Skill has genuine value. If the answer is no → it's compressing what Claude already knows.

The best Skills are compressed expert brains — they take a designer's 10 years of aesthetic accumulation and compress it into 43 lines, or a document expert's operational experience into a 200-line decision tree.

What gets compressed must be things Claude doesn't have. Otherwise, it's garbage compression.

---

Self-Evaluation Note

This Skill (skill-judge) should itself pass evaluation:

• Knowledge Delta: Provides specific evaluation criteria Claude wouldn't generate on its own
• Mindset: Shapes how to think about Skill quality, not just checklist items
• Anti-Patterns: "NEVER Do When Evaluating" section with specific don'ts
• Specification: Valid frontmatter with comprehensive description
• Progressive Disclosure: Self-contained, no external references needed
• Freedom: Medium freedom appropriate for evaluation task
• Pattern: Follows Tool pattern with decision frameworks
• Usability: Clear protocol, report template, quick reference

Evaluate this Skill against itself as a calibration exercise.