Skill Linter

When to Use

• Adding new skills to the marketplace

• Reviewing skill PRs

• Running quality gates before merge

• Checking existing skills for compliance

Validation Rules

Required Frontmatter

Field Constraints

name

1-64 chars, lowercase alphanumeric + hyphens, no leading/trailing/consecutive hyphens, must match parent directory name. Cannot contain "claude" or "anthropic" (reserved).

description

1-1024 chars, non-empty, should include keywords for discoverability. No XML angle brackets (< >).

Optional Frontmatter

Field Constraints

compatibility

1-500 chars, environment requirements

metadata

Key-value pairs (string values only)

allowed-tools

Space-delimited tool list (FAIL on commas or array syntax)

Structure Requirements

Rule Requirement

Directory name Must match name field exactly

SKILL.md Required, must exist

README.md Must NOT exist inside skill folder (docs go in SKILL.md or references/)

Line limit Max 500 lines in SKILL.md

Subdirectories Only scripts/ , references/ , assets/ allowed

Content Quality Rules

Rule Requirement

No ASCII art Box-drawing characters (─│┌┐└┘├┤┬┴┼), arrows (↑↓←→↔), and decorative diagrams waste tokens. LLMs tokenize character-by-character, not visually. Use plain lists or tables instead.

No decorative quotes Inspirational quotes or attributions ("As X said...") have no functional value for LLM execution.

No persona statements "You are an expert..." wastes tokens. Use Audience: / Goal: framing instead.

Functional content only Every line should improve LLM behavior. Ask: "Does this help Claude execute better?"

Audience/Goal Framing (Required)

Replace persona roleplay with audience-focused framing:

❌ Bad (persona):

You are an expert software engineer with deep expertise in testing. Your role is to analyze code and generate thorough test coverage.

✅ Good (audience/goal):

Audience: Developers needing test coverage for new or changed code.

Goal: Generate comprehensive tests based on specified test type and framework.

Rationale: "Explain X for audience Y" yields better-tailored outputs than "Act as persona Z".

ASCII Art Detection Pattern:

[─│┌┐└┘├┤┬┴┼╭╮╯╰═║╔╗╚╝╠╣╦╩╬↑↓←→↔⇒⇐⇔▲▼◄►]{3,}

Files matching this pattern should be flagged for review.

Description Routing Quality (WARN)

Skill descriptions are routing logic, not documentation. They answer:

• When to use this skill (trigger conditions)

• When NOT to use (negative routing)

• What outputs to expect (success criteria)

Quality Level Example

Good "Use when implementing Stimulus controllers. Not for React components. Outputs controller with targets and actions."

Adequate "Use when working with Stimulus controllers in Rails applications."

Poor "Stimulus controller development skill."

Anti-pattern "Comprehensive, powerful toolkit for building cutting-edge Stimulus controllers."

Templates inside skills are encouraged — loaded only on trigger, essentially free tokens.

Skill Disambiguation (INFO)

When multiple skills cover similar domains, include negative routing:

Example (minitest vs rspec):

Skill Description Routing

minitest-coder

"Use when writing Minitest tests. Not for RSpec — use rspec-coder instead."

rspec-coder

"Use when writing RSpec tests. Not for Minitest — use minitest-coder instead."

"Don't use when..." is as important as "Use when..." for routing accuracy.

Name Pattern

^[a-z][a-z0-9](-[a-z0-9]+)$

Valid: my-skill , skill1 , api-v2-handler

Invalid: -skill , skill- , my--skill , MySkill , my_skill

Command/Agent Invocation Patterns

Skills should primarily provide knowledge, not orchestration. If invocations are needed:

Pattern Status Use Instead

Skill("command", args: "...")

❌ Deprecated /command args

SlashCommand("command", ...)

❌ Deprecated /command args

Task(subagent_type="agent", ...)

✅ Correct (no change)

✅ Preferred command invocation:

/majestic:config tech_stack generic /majestic-engineer:tdd-workflow /majestic-ralph:start "task" --max-iterations 50

❌ Deprecated patterns:

Skill("config-reader", args: "tech_stack generic") SlashCommand("majestic:build-task", args: "...")

Note: Agent invocation via Task() is correct - there is no @agent syntax.

Usage

Validate Single Skill

./scripts/validate-skill.sh path/to/skill-name

Validate All Marketplace Skills

for skill in plugins//skills//; do ./scripts/validate-skill.sh "$skill" done

CI Integration

Add to pre-commit hook or CI pipeline:

• name: Lint Skills run: | for skill in plugins//skills//; do .claude/skills/skill-linter/scripts/validate-skill.sh "$skill" || exit 1 done

Validation Script

The linter script at scripts/validate-skill.sh performs these checks:

• Directory exists with SKILL.md file

• Frontmatter present with YAML delimiters

• Name field valid - pattern, length, matches directory

• Reserved words - name cannot contain "claude" or "anthropic" (FAIL)

• Description field valid - present, length constraints

• XML tags in frontmatter - no angle brackets < > (security, FAIL)

• No README.md - inside skill folder (WARN)

• Optional fields valid - if present, match constraints

• Line count - under 500 lines

• Subdirectory names - only allowed directories

• No ASCII art - box-drawing characters outside fenced code blocks (WARN)

• No persona statements - "You are a/an/the" outside fenced code blocks (FAIL)

• Description routing quality - routing keywords present (WARN)

• Marketing copy detection - buzzwords in description (WARN)

Error Codes

Code Meaning

0 All validations passed

1 Missing SKILL.md

2 Invalid frontmatter

3 Name validation failed

4 Description validation failed

5 Optional field validation failed

6 Line limit exceeded

7 Invalid subdirectory

8 ASCII art detected (warning)

9 Persona statement detected

10 Description routing quality (warning)

11 Marketing copy detected (warning)

Example Output

Validating: plugins/majestic-tools/skills/brainstorming

[PASS] SKILL.md exists [PASS] Frontmatter present [PASS] Name 'brainstorming' valid (12 chars) [PASS] Name matches directory [PASS] Description valid (156 chars) [PASS] Line count: 87/500 [PASS] Subdirectories valid [PASS] No ASCII art outside code blocks [PASS] No persona statements [PASS] Description has routing keywords [PASS] No marketing copy in description

Result: ALL CHECKS PASSED

Spec Reference

Based on agentskills.io/specification:

• Progressive disclosure - Metadata ~100 tokens at startup, full content <5000 tokens when activated

• Reference files - Keep one level deep from SKILL.md

• Token budget - Main SKILL.md recommended under 500 lines