Skills Expert

Provide expert guidance on Claude Code skills development, validation, and troubleshooting.

Overview

This skill transforms Claude into a Claude Skills architecture expert. Use when:

• Creating new skills
• Debugging skill discovery/invocation issues
• Validating YAML frontmatter
• Understanding tool permissions
• Optimizing skill descriptions for selection

Prerequisites

• Access to skill files in .claude/skills/ or skills-pack/.claude/skills/
• Understanding of target skill's purpose

Claude Skills Architecture

YAML Frontmatter Fields

Required Fields

| Field | Max Length | Purpose | |-------|------------|---------| | name | 64 chars | Identifier used as command in Skill tool invocation | | description | 1024 chars | Primary signal Claude uses to decide when to invoke skill |

Optional Fields

| Field | Type | Purpose | |-------|------|---------| | allowed-tools | CSV string | Pre-approved tools during skill execution | | model | string | Override model ("inherit" or model ID) | | version | string | Semantic versioning (e.g., "1.0.0") | | license | string | License metadata | | mode | boolean | true = appears in Mode Commands section | | disable-model-invocation | boolean | true = requires manual /skill-name |

Undocumented but Functional

| Field | Behavior | |-------|----------| | when_to_use | Appends to description with hyphen separator |

Critical Rule: Skill MUST have description OR when_to_use or it's filtered out.

Directory Structure

skill-name/
├── SKILL.md              # Core prompt + frontmatter (required)
├── scripts/              # Python/Bash executables
├── references/           # Documentation loaded via Read tool
└── assets/               # Templates, static files (referenced by path)

{baseDir} Variable

Template variable resolving to skill installation directory:

## Resources
- Skill standard: `{baseDir}/references/skill-standard.md`

Never hardcode absolute paths.

Tool Permission Scoping

Permissions in allowed-tools are scoped to skill execution only:

# Granular scoping examples
allowed-tools: "Bash(git:*),Read,Grep"     # Only git commands
allowed-tools: "Bash(npm:*),Read,Write"    # Only npm commands
allowed-tools: "Read,Glob,Grep"            # Read-only audit

Skill Selection Mechanism

Claude decides skill invocation through pure language understanding:

1. Claude receives Skill tool with <available_skills> section
2. Each skill formatted as: "skill-name": description
3. Claude's transformer matches user intent to descriptions
4. Claude invokes Skill tool with matching command

No embeddings, regex, or classifiers - pure LLM reasoning.

Execution Flow

1. User request matches skill description
2. Claude invokes Skill tool with command: "skill-name"
3. Two messages injected:
   • Visible: <command-message>The "skill-name" skill is loading</command-message>
   • Hidden: Full SKILL.md content (isMeta: true)
4. Tool permissions scoped to allowed-tools
5. Skill prompt expands into conversation context

Token Budget

• Default skill description budget: 15,000 characters
• Typical skill prompt: 500-5,000 words

Discovery Priority

Skills loaded in order (later overrides earlier):

1. User settings (~/.config/claude/skills/)
2. Project settings (.claude/skills/)
3. Plugin-provided skills
4. Built-in skills

Critical Gotchas

1. Skills are NOT concurrency-safe - Multiple simultaneous skill invocations cause context conflicts
2. Skills don't live in system prompts - They're in the tools array as part of Skill meta-tool
3. when_to_use is undocumented/experimental - Safer to rely on detailed description
4. Hardcoded paths break portability - Always use {baseDir}

Three-Stage Execution Pipeline

1. VALIDATION
   - Syntax checking
   - Skill existence verification
   - Frontmatter parsing

2. PERMISSION EVALUATION
   - Deny rules checked first (blocking patterns)
   - Allow rules checked second (pre-approved)
   - Default: prompt user for approval

3. LOADING & INJECTION
   - SKILL.md content loaded
   - Two messages injected (visible + hidden)
   - Context modifier applied
   - Tool permissions scoped

Where Skills Actually Live

Skills are NOT in system prompts. They're bundled in the tools array:

tools: [
  { name: "Read", ... },
  { name: "Write", ... },
  {
    name: "Skill",           // Meta-tool
    inputSchema: { command: string },
    prompt: "<available_skills>..." // Dynamic, contains all skill descriptions
  }
]

This enables dynamic loading without system prompt manipulation.

Instructions

When Creating Skills

1. Write action-oriented descriptions that explicitly state use cases
2. Keep descriptions under 1024 chars - be concise
3. Use minimal tool permissions - principle of least privilege
4. Use {baseDir} for all path references
5. Include all four directories (scripts/, references/, assets/) even if empty
6. Test discovery by asking Claude "what skills do you have?"

When Debugging Skills

1. Check YAML frontmatter syntax (no tabs, proper quoting)
2. Verify description or when_to_use exists
3. Check skill directory is in correct location
4. Verify SKILL.md filename (case-insensitive)
5. Check for disable-model-invocation: true blocking auto-discovery

When Validating Compliance

# Nixtla Standard Checklist
- [ ] name: lowercase + hyphens, matches folder name
- [ ] description: action-oriented, <1024 chars
- [ ] version: semver format
- [ ] allowed-tools: minimal necessary
- [ ] NO deprecated fields (author, priority, audience)
- [ ] mode: true ONLY for mode skills
- [ ] disable-model-invocation: true ONLY for infra/dangerous skills

Output

• Validated SKILL.md frontmatter
• Corrected skill configurations
• Debugging recommendations
• Skill architecture explanations

Error Handling

| Issue | Solution | |-------|----------| | Skill not discovered | Check description/when_to_use exists | | Tools not working | Verify allowed-tools syntax | | Wrong skill invoked | Improve description specificity | | Skill filtered out | Add required frontmatter fields |

Examples

• Diagnose why a skill does not auto-trigger, then propose an updated description that includes both “Use when …” and “Trigger with …”.
• Audit a skill for minimal allowed-tools and scoped Bash(...) usage.

Resources

• Project validator: 004-scripts/validate_skills_v2.py
• Project skills: 003-skills/.claude/skills/

Examples

Example 1: Minimal Skill

---
name: my-skill
description: Generate reports from CSV data. Use when user has CSV files and needs analysis.
allowed-tools: "Read,Write,Glob"
version: "1.0.0"
---

Example 2: Mode Skill

---
name: expert-mode
description: Transform Claude into domain expert for extended session.
mode: true
allowed-tools: "Read,Write,Glob,Grep,Edit,Bash"
version: "1.0.0"
---

Example 3: Infrastructure Skill

---
name: deploy-skill
description: Deploy application to production. Dangerous - requires explicit invocation.
disable-model-invocation: true
allowed-tools: "Bash(deploy:*),Read,Glob"
version: "1.0.0"
---

Resources

• Skill standard: {baseDir}/references/skill-standard.md
• Deep dive source: https://leehanchung.github.io/blogs/2025/10/26/claude-skills-deep-dive/