@vibe-agent-toolkit/runtime-claude-skills

Build, validate, and import Claude Skills for Claude Desktop, Claude Code, and VAT agents.

Overview

This package provides runtime support for Claude Skills including:

• Validation - Audit Claude Skills for quality and compatibility
• Import/Export - Convert between SKILL.md and agent.yaml formats
• Building - Package VAT agents as Claude Skills
• Frontmatter Parsing - Extract and validate Claude Skills frontmatter

Features

• Skill Validation - Comprehensive validation against Agent Skills specification
• Format Conversion - Import SKILL.md to agent.yaml (and vice versa)
• Build Pipeline - Generate SKILL.md from agent manifests
• Schema Validation - Zod-based frontmatter validation with TypeScript types
• Link Integrity - Validate all markdown links and references
• Console Compatibility - Detect tool usage incompatible with console mode

Installation

npm install @vibe-agent-toolkit/runtime-claude-skills

Quick Start

Validate a Claude Skill

import { validateSkill } from '@vibe-agent-toolkit/runtime-claude-skills';

const result = await validateSkill({
  skillPath: './my-skill/SKILL.md',
});

if (result.status === 'error') {
  console.error('Validation failed:');
  for (const issue of result.issues) {
    console.error(`  [${issue.code}] ${issue.message}`);
  }
}

Import SKILL.md to agent.yaml

import { importSkillToAgent } from '@vibe-agent-toolkit/runtime-claude-skills';

const result = await importSkillToAgent({
  skillPath: './my-skill/SKILL.md',
  outputPath: './my-agent/agent.yaml', // Optional
  force: false, // Optional
});

if (result.success) {
  console.log(`Imported to: ${result.agentPath}`);
} else {
  console.error(`Import failed: ${result.error}`);
}

Build Claude Skill from VAT Agent

import { buildClaudeSkill } from '@vibe-agent-toolkit/runtime-claude-skills';

await buildClaudeSkill({
  agentPath: './my-agent',
  outputPath: './dist/skills/my-agent',
});

API Reference

validateSkill(options): Promise

Validate a Claude Skill (SKILL.md) for quality and compatibility.

Options:

interface ValidateOptions {
  skillPath: string;         // Path to SKILL.md
  rootDir?: string;          // Root directory for resolving links
  isVATGenerated?: boolean;  // Treat as VAT-generated (stricter validation)
}

Returns:

interface ValidationResult {
  path: string;                       // Path to skill file
  type: 'claude-skill' | 'vat-agent'; // Detected type
  status: 'success' | 'warning' | 'error';
  summary: string;                    // Human-readable summary
  issues: ValidationIssue[];          // All validation issues
  metadata?: {                        // Extracted metadata
    name?: string;
    description?: string;
    lineCount?: number;
  };
}

interface ValidationIssue {
  severity: 'error' | 'warning' | 'info';
  code: IssueCode;              // Machine-readable code
  message: string;              // Human-readable message
  location?: string;            // File:line location
  fix?: string;                 // Suggested fix
}

Example:

const result = await validateSkill({
  skillPath: './my-skill/SKILL.md',
  rootDir: './my-skill',
});

console.log(`Status: ${result.status}`);
console.log(`Summary: ${result.summary}`);

// Check for specific error types
const nameErrors = result.issues.filter(
  i => i.code === 'SKILL_NAME_INVALID'
);

if (nameErrors.length > 0) {
  console.error('Name validation failed');
}

Validation Rules

The validator checks for:

Critical Errors (Blocking)

Frontmatter Errors:

• SKILL_MISSING_FRONTMATTER - No YAML frontmatter found
• SKILL_MISSING_NAME - Required "name" field missing
• SKILL_MISSING_DESCRIPTION - Required "description" field missing
• SKILL_NAME_INVALID - Name doesn't match pattern: ^[a-z0-9]+(-[a-z0-9]+)*$
• SKILL_NAME_RESERVED_WORD - Name contains "claude" or "anthropic"
• SKILL_NAME_XML_TAGS - Name contains < or > characters
• SKILL_DESCRIPTION_TOO_LONG - Description exceeds 1024 characters
• SKILL_DESCRIPTION_EMPTY - Description is empty or whitespace-only
• SKILL_DESCRIPTION_XML_TAGS - Description contains < or > characters

Link Errors:

• LINK_INTEGRITY_BROKEN - Link points to non-existent file
• PATH_STYLE_WINDOWS - Link uses Windows backslashes ()

Warnings (Non-Blocking)

• SKILL_TOO_LONG - Skill exceeds 5000 lines
• SKILL_CONSOLE_INCOMPATIBLE - References Write, Edit, Bash, or NotebookEdit tools

See Best Practices Guide for detailed guidance.

importSkillToAgent(options): Promise

Convert a Claude Skill (SKILL.md) to VAT agent format (agent.yaml).

Options:

interface ImportOptions {
  skillPath: string;      // Path to SKILL.md
  outputPath?: string;    // Custom output path (default: same dir as SKILL.md)
  force?: boolean;        // Overwrite existing agent.yaml (default: false)
}

Returns:

type ImportResult =
  | { success: true; agentPath: string }
  | { success: false; error: string };

Example:

// Basic import
const result = await importSkillToAgent({
  skillPath: './my-skill/SKILL.md',
});

// Custom output path
const result = await importSkillToAgent({
  skillPath: './my-skill/SKILL.md',
  outputPath: './agents/my-agent/agent.yaml',
});

// Force overwrite
const result = await importSkillToAgent({
  skillPath: './my-skill/SKILL.md',
  force: true,
});

if (!result.success) {
  console.error(`Import failed: ${result.error}`);
  process.exit(1);
}

What Gets Converted:

| SKILL.md Field | agent.yaml Field | Notes | |----------------|------------------|-------| | name | metadata.name | Required | | description | metadata.description | Required | | metadata.version | metadata.version | Defaults to "0.1.0" | | metadata.tags | metadata.tags | Optional | | license | metadata.license | Optional | | compatibility | spec.compatibility | Optional |

The generated agent.yaml will have:

metadata:
  name: skill-name
  description: Skill description
  version: 1.0.0
spec:
  runtime: claude-skills

buildClaudeSkill(options): Promise

Build a Claude Skill from a VAT agent manifest.

Options:

interface BuildOptions {
  agentPath: string;    // Path to agent directory or manifest
  outputPath: string;   // Where to write skill bundle
}

Returns:

interface BuildResult {
  outputPath: string;   // Path to generated SKILL.md
  metadata: {
    name: string;
    description: string;
    version: string;
  };
}

Example:

const result = await buildClaudeSkill({
  agentPath: './my-agent',
  outputPath: './dist/skills/my-agent',
});

console.log(`Built skill: ${result.outputPath}`);
console.log(`Version: ${result.metadata.version}`);

Frontmatter Schemas

ClaudeSkillFrontmatterSchema

Strict schema for console-compatible Claude Skills:

import { ClaudeSkillFrontmatterSchema } from '@vibe-agent-toolkit/runtime-claude-skills';
import { z } from 'zod';

const frontmatter = {
  name: 'my-skill',
  description: 'Does something useful',
  compatibility: 'Requires Node.js 18+',
};

const result = ClaudeSkillFrontmatterSchema.safeParse(frontmatter);
if (result.success) {
  // TypeScript type: ClaudeSkillFrontmatter
  const validated = result.data;
}

Required Fields:

• name - Skill identifier (lowercase, hyphens, max 64 chars)
• description - What skill does (max 1024 chars)

Optional Fields:

• license - License identifier
• compatibility - Environment requirements (max 500 chars)
• metadata - Additional properties (Record<string, string>)
• allowed-tools - Pre-approved tools (experimental)

VATClaudeSkillFrontmatterSchema

Extended schema for VAT-generated skills:

import { VATClaudeSkillFrontmatterSchema } from '@vibe-agent-toolkit/runtime-claude-skills';

const frontmatter = {
  name: 'my-skill',
  description: 'Does something useful',
  metadata: {
    version: '1.0.0',   // Required for VAT skills
    tags: ['utility', 'data'],
  },
};

const result = VATClaudeSkillFrontmatterSchema.safeParse(frontmatter);

Additional Requirements:

• metadata.version - Semantic version (required for VAT)

parseFrontmatter(content): ParseResult

Parse YAML frontmatter from SKILL.md content.

Example:

import { parseFrontmatter } from '@vibe-agent-toolkit/runtime-claude-skills';
import * as fs from 'node:fs';

const content = fs.readFileSync('./SKILL.md', 'utf-8');
const result = parseFrontmatter(content);

if (result.success) {
  console.log('Frontmatter:', result.frontmatter);
  console.log('Body:', result.body);
} else {
  console.error('Parse error:', result.error);
}

Usage Examples

Validate Before Import

import {
  validateSkill,
  importSkillToAgent
} from '@vibe-agent-toolkit/runtime-claude-skills';

// Validate first
const validation = await validateSkill({
  skillPath: './my-skill/SKILL.md',
});

if (validation.status === 'error') {
  console.error('Validation failed, cannot import');
  process.exit(1);
}

// Import if validation passes
const importResult = await importSkillToAgent({
  skillPath: './my-skill/SKILL.md',
});

if (importResult.success) {
  console.log(`Imported to: ${importResult.agentPath}`);
}

Batch Validation

import { validateSkill } from '@vibe-agent-toolkit/runtime-claude-skills';
import * as fs from 'node:fs/promises';
import * as path from 'node:path';

async function validateAllSkills(dir: string) {
  const entries = await fs.readdir(dir, { withFileTypes: true });
  const results = [];

  for (const entry of entries) {
    if (entry.isDirectory()) {
      const skillPath = path.join(dir, entry.name, 'SKILL.md');
      try {
        const result = await validateSkill({ skillPath });
        results.push({ skill: entry.name, result });
      } catch (error) {
        console.error(`Error validating ${entry.name}:`, error);
      }
    }
  }

  return results;
}

const results = await validateAllSkills('./skills');
const failed = results.filter(r => r.result.status === 'error');

console.log(`Validated ${results.length} skills`);
console.log(`Failed: ${failed.length}`);

Custom Validation Logic

import { validateSkill } from '@vibe-agent-toolkit/runtime-claude-skills';

const result = await validateSkill({
  skillPath: './my-skill/SKILL.md',
});

// Filter by severity
const errors = result.issues.filter(i => i.severity === 'error');
const warnings = result.issues.filter(i => i.severity === 'warning');

// Group by error code
const byCode = result.issues.reduce((acc, issue) => {
  if (!acc[issue.code]) acc[issue.code] = [];
  acc[issue.code].push(issue);
  return acc;
}, {} as Record<string, typeof result.issues>);

// Check for specific issues
const hasLinkErrors = result.issues.some(
  i => i.code === 'LINK_INTEGRITY_BROKEN'
);

if (hasLinkErrors) {
  console.error('Fix broken links before proceeding');
}

Integration

CLI Integration

This package powers these CLI commands:

• vat agent audit - Validates skills using validateSkill()
• vat agent import - Imports skills using importSkillToAgent()

See CLI Documentation for command usage.

CI/CD Integration

// ci-validate.ts
import { validateSkill } from '@vibe-agent-toolkit/runtime-claude-skills';
import * as fs from 'node:fs/promises';

async function validateInCI() {
  const files = await fs.readdir('./skills', { recursive: true });
  const skills = files.filter(f => f.endsWith('SKILL.md'));

  let hasErrors = false;

  for (const skill of skills) {
    const result = await validateSkill({
      skillPath: `./skills/${skill}`,
    });

    if (result.status === 'error') {
      console.error(`❌ ${skill}:`);
      for (const issue of result.issues) {
        if (issue.severity === 'error') {
          console.error(`  ${issue.message}`);
        }
      }
      hasErrors = true;
    }
  }

  if (hasErrors) {
    process.exit(1);
  }

  console.log('✅ All skills validated successfully');
}

validateInCI();

TypeScript Types

The package exports all types for TypeScript users:

import type {
  ValidateOptions,
  ValidationResult,
  ValidationIssue,
  IssueCode,
  IssueSeverity,
  ClaudeSkillFrontmatter,
  VATClaudeSkillFrontmatter,
  ImportOptions,
  ImportResult,
  BuildOptions,
  BuildResult,
} from '@vibe-agent-toolkit/runtime-claude-skills';

Error Handling

All async functions use standard Promise rejection for errors:

try {
  const result = await validateSkill({
    skillPath: './invalid-path/SKILL.md',
  });
} catch (error) {
  if (error instanceof Error) {
    console.error(`Validation error: ${error.message}`);
  }
}

Validation errors are returned in the ValidationResult object, not thrown:

const result = await validateSkill({ skillPath: './skill/SKILL.md' });

// result.status will be 'error', not thrown
if (result.status === 'error') {
  // Handle validation failures
}

Performance

Validation Performance

Validation is optimized for:

• Fast parsing - Efficient YAML frontmatter extraction
• Cached git checks - Gitignore status cached per directory
• Parallel validation - Validate multiple skills concurrently

Typical performance:

• Single skill: ~10-50ms
• 10 skills: ~100-200ms (parallel)
• 100 skills: ~1-2s (parallel)

Memory Usage

Memory usage is proportional to skill size:

• Small skill (<100KB): ~2MB
• Medium skill (500KB): ~5MB
• Large skill (2MB): ~10MB

For large-scale validation, validate in batches to control memory.

Cross-Platform Compatibility

The package is fully cross-platform:

• Path handling - Uses node:path for Windows/Unix compatibility
• Line endings - Handles CRLF and LF correctly
• File system - Works with case-sensitive and case-insensitive file systems

Tested on:

• Windows 10/11
• macOS (Apple Silicon and Intel)
• Linux (Ubuntu, Debian, Alpine)

Related Packages

• @vibe-agent-toolkit/discovery - File discovery for skills
• @vibe-agent-toolkit/cli - CLI commands using this package
• @vibe-agent-toolkit/utils - Shared utilities

Related Documentation

• Audit Command - CLI validation command
• Import Command - CLI import command
• Claude Skills Best Practices - Comprehensive guide
• Agent Skills Specification - Official spec

Testing

The package includes comprehensive tests:

# Unit tests
bun run test:unit

# Integration tests
bun run test:integration

# Watch mode
bun run test:watch

License

MIT