Agent Skills

Agent Skills is my shared repo of reusable skills, commands, and config for AI coding agents. I wanted one place to keep the instructions and scripts I actually rely on, then make them available across Claude Code, Codex, and Pi. It is open source, but the main purpose is to support my own workflow with tools I can inspect and change directly.

Problem

Installing agent skills from npm felt like the wrong security model for something that can shape local workflows and run scripts. I would rather review plain files in a repo I control than depend on package installs, opaque updates, and code that is harder to tweak in the middle of real work.

The other problem was duplication. I was switching between different agent harnesses, re-creating the same setup, and re-explaining repeated workflows in prompts instead of giving the agent a reusable layer it could discover on its own.

Solution

I keep broadly useful skills in one repo I own and edit directly, then symlink them into each agent’s skills directory. That gives me one source of truth across harnesses while still leaving room for project-specific skills inside individual repos.

A lot of the repo came from seeing useful ideas in other people’s skills, then adapting them into versions that better fit how I work. That makes the repo both a toolkit and a place to iterate: if I notice a failure mode, I can update the skill or add a helper script immediately.

What I built

The repo combines a few different layers:

• reusable reference skills for specific stacks, tools, and workflows
• command-style skills for repeated tasks like review, implementation, and CI checks
• small config helpers for things like notifications and statusline behavior
• a setup script that links one source repo into multiple agent environments

Technical architecture

• Skills live as plain files under skills/, one directory per skill.
• They use the shared SKILL.md format with YAML frontmatter.
• setup.sh symlinks them into ~/.claude/skills, ~/.agents/skills, and ~/.pi/agent/skills.
• Because they are symlinked, any edit in the repo is immediately available everywhere.

Product decisions

• Prefer locally owned files over npm-installed skills.
• Keep reusable skills here; keep project-specific skills in each project.
• Adapt useful ideas from other people’s skills instead of treating them as fixed packages.
• Add a skill only when a workflow repeats or the agent keeps making the same mistake.
• Avoid adding context the model already knows.
• Keep the repo focused instead of turning it into a giant catalog.

Why it has been useful

Skills make repeated workflows more reliable. They reduce copy-paste, make useful context discoverable to the agent, and keep my tooling more consistent across different harnesses. That consistency also makes it easier to compare models on a more even baseline, since more of the surrounding setup stays the same.

Current status

This is an actively evolving personal toolkit that I use in real work. It is public so other people can inspect it, copy it, and adapt it for their own setup.