agentskills-io

Review·Scanned 2/18/2026

Provides templates, documentation, and runnable helper scripts for creating and validating Agent Skills. Contains executable shell commands (e.g., uvx --from git+https://github.com/agentskills/agentskills#subdirectory=skills-ref) and scripts that modify files such as plugins/${plugin}/.claude-plugin/plugin.json, enabling local execution and external network fetches.

from clawhub.ai·v1.0·24.7 KB·0 installs
Scanned from 2.5.0 at 7224886 · Transparency log ↗
$ vett add clawhub.ai/killerapp/agentskills-ioReview findings below

Agent Skills (agentskills.io)

Create portable skills for AI agents. Works with Claude Code, Cursor, GitHub Copilot, OpenAI integrations, VS Code (symlinks enable sharing across tools).

Resources

Structure

skill-name/
├── SKILL.md          # Required (frontmatter + instructions, <5000 tokens activation)
├── scripts/          # Optional: executable code
├── references/       # Optional: detailed docs
└── assets/           # Optional: templates, static files

Rules: Dir name = frontmatter name:. Only 3 subdirs. SKILL.md <500 lines. ~100 tokens for discovery (name+desc).

Frontmatter

Required

  • name: 1-64 chars, lowercase alphanumeric-hyphens (^[a-z0-9]+(-[a-z0-9]+)*$)
  • description: 1-1024 chars, include "Use when..." (discovery budget: ~100 tokens)

Optional

  • license: SPDX identifier (Apache-2.0, MIT) | compatibility: Environment reqs (<500 chars)
  • metadata: Key-value pairs (author, version, tags) | allowed-tools: Space-delimited tool list

Validation

# Install permanently (vs ephemeral uvx)
uv tool install git+https://github.com/agentskills/agentskills#subdirectory=skills-ref
# Or use uvx for one-shot validation
uvx --from git+https://github.com/agentskills/agentskills#subdirectory=skills-ref skills-ref validate ./skill
CommandDescription
skills-ref validate <path>Check structure, frontmatter, token budgets
skills-ref read-properties <path>Extract metadata
skills-ref to-prompt <path>Generate prompt format

Writing Rules

  • Imperative language: "Check: command" not "You might want to..."
  • Concrete examples with expected output; handle common errors with solutions
  • Progressive disclosure: core in SKILL.md (<5000 tokens), details in references/

Common Errors

ErrorFix
Invalid nameLowercase alphanumeric-hyphens only
Missing descriptionAdd description: field with "Use when..."
Description too long<1024 chars, move details to body
Invalid YAMLCheck indentation, quote special chars
Missing SKILL.mdFilename must be exactly SKILL.md
Dir name mismatchDirectory name must match name: field

Quick Workflow

  1. Create: mkdir skill-name && touch skill-name/SKILL.md
  2. Add frontmatter (name, description with "Use when...")
  3. Write instructions (bullets, not prose); validate: skills-ref validate ./skill-name
  4. Test with AI agent, iterate; add LICENSE, push to repository

Plugin Structure (Claude Code)

plugin-name/
├── .claude-plugin/plugin.json
├── README.md, LICENSE, CHANGELOG.md  # CHANGELOG.md tracks versions
├── skills/skill-name/SKILL.md
├── agents/     # Optional: subagents (.md files)
└── examples/   # Optional: full demo projects

Distinctions: Plugin examples/ = runnable projects. Skill assets/ = static resources only.

Batch Validation & Versioning

bash scripts/validate-skills-repo.sh     # Validate all skills in repo
bash scripts/bump-changed-plugins.sh     # Auto-bump only changed plugins (semver)

Minimal Example

---
name: example-skill
description: Brief description. Use when doing X.
---
# Example Skill
## Prerequisites
- Required tools
## Instructions
1. First step: `command`
2. Second step with example
## Troubleshooting
**Error**: Message → **Fix**: Solution

Symlink Sharing

Share skills across Claude Code, Cursor, VS Code: ln -s /path/to/skills ~/.cursor/skills

References

  • specification.md - Full YAML schema, token budgets
  • examples.md - Complete examples across platforms
  • validation.md - Error troubleshooting
  • best-practices.md - Advanced patterns, symlink setup