Reusable skills for AI coding agents. Works with pi, Claude Code, Cursor, and any agent that supports the Agent Skills standard.
| Category | Skill | Description |
|---|---|---|
| ☁️ Google Cloud | gcloud | GCP CLI with agent safety model — hub + 7 reference files |
| ⚙️ Process | pm2 | Process management — keep services alive, auto-restart, monitoring, ecosystem configs |
| 🚀 Release | pre-release | Pre-release checklist + AI-written changesets via @changesets/cli |
| 🔍 Search | web-search | Web search + content extraction via ddgs — no API keys |
| 🎬 Terminal | vhs | Record terminal sessions as GIF/MP4 with VHS |
npx skills add marcfargas/skillsInstalls to Claude Code, Cursor, Copilot, Amp, Cline, Windsurf, Gemini CLI, and 30+ more agents automatically.
pi install npm:@marcfargas/skillsOr add to ~/.pi/agent/settings.json:
{
"packages": ["npm:@marcfargas/skills"]
}Copy the skill directory into your agent's skill folder:
cp -r google-cloud/gcloud ~/.claude/skills/gcloudEvery skill is reviewed by 3+ models (Claude, Gemini, GPT) before publishing — structure, agent usability, safety, and real-world scenario testing. If an agent can misinterpret an instruction, we find out before you do.
Every operation is classified: READ / WRITE / DESTRUCTIVE / EXPENSIVE / FORBIDDEN. Destructive and expensive operations are gated — the agent must confirm before executing, and costs are flagged upfront.
Skills use a hub + spoke architecture. The SKILL.md hub is ~140 lines — just enough to match the right skill and know what's available. Detailed per-topic reference files are loaded on demand, keeping your context window lean.
- Agent-native —
--format=jsoneverywhere, idempotent patterns, structured error handling - Portable — no hardcoded paths, no personal config, works on any machine
- Spec-compliant — validated against the Agent Skills specification using skills-ref in CI
- Continuous validation —
agentskills validateon every push (validate.yml), pre-release checklist with AI-written changesets, npm Trusted Publishing with provenance
skills/
├── google-cloud/
│ └── gcloud/ # 8 files, ~1100 lines total
├── process/
│ └── pm2/ # 1 file
├── release/
│ └── pre-release/ # 1 file
├── search/
│ └── web-search/ # SKILL.md + search.js + content.js
├── terminal/
│ └── vhs/ # 1 file
└── README.md
Some skills live in their own repositories — install them directly or via their npm packages:
| Skill | Description | Install |
|---|---|---|
| go-easy | Gmail, Drive, Calendar for AI agents — npx go-gmail, npx go-drive, npx go-calendar |
npx skills add marcfargas/go-easy |
| holdpty | Detached PTY sessions — launch, attach, view, record terminal processes | npx skills add marcfargas/holdpty |
| odoo | Odoo ERP integration — connect, introspect, automate | npx skills add marcfargas/odoo-toolbox |
Skills follow the Agent Skills specification. Requirements:
SKILL.mdwith YAML frontmatter (name,description)namematches parent directory (kebab-case, max 64 chars)descriptionpresent (max 1024 chars)- No hardcoded paths or credentials
- Destructive operations clearly marked
Use skills-ref (Python — the official reference implementation from the spec authors) to validate skills locally:
pip install skills-ref
# Validate a skill directory
agentskills validate path/to/skill
# Read parsed properties as JSON
agentskills read-properties path/to/skill
# Generate <available_skills> XML prompt block
agentskills to-prompt path/to/skill-a path/to/skill-bCI runs agentskills validate on every push — see .github/workflows/validate.yml.
Building high-quality, multi-model-reviewed agent skills takes serious token budget. If these skills save you time, consider sponsoring:
Code (scripts, tooling): MIT
Skill content (**/SKILL.md and reference docs): CC0 1.0 Universal — public domain, no attribution required.
Why CC0 for skills: Skill docs are consumed by AI agents and freely incorporated into any workflow. Attribution requirements create friction in agent contexts where provenance tracking is impractical.