If you’ve spent any time building with AI agents, you’ve hit the same wall: agents are powerful but unstructured. They can chat, they can code, but making them reliably execute complex workflows? That’s still more art than engineering.
Enter davidondrej/skills — an open-source collection of reusable agent skills that turns AI agents from chatty generalists into focused operators.
What Are Agent Skills?
An agent skill is a packaged workflow — a SKILL.md file that tells an AI agent exactly what to do, when to trigger, and how to execute. Think of it as a recipe card that an agent can pick up and follow when the task calls for it.
The concept is simple but powerful: instead of prompting an agent from scratch every time, you give it a library of skills it can load on demand. Each skill is self-contained, versioned, and composable.
The Repo at a Glance
The repo organizes skills into five categories:
Agent Orchestration
Running, scheduling, delegating, and coordinating AI agents.
- agent-self-scheduling — Make an agent run on a schedule, loop, or interval (cron, heartbeats, recurring checks). Covers external clocks vs built-in schedulers.
- cmux — Control the cmux macOS terminal app (CLI + socket API) for workspaces, panes, surfaces, browser automation, notifications, and hooks.
- codex-subagent — Delegate work to Codex as a sub-agent.
- fable-review / fable-safe-prompt — Review and safely prompt with Fable (an agent framework).
- git-worktree — Manage git worktrees for parallel agent work.
This is where the meta lives. Skills that orchestrate other agents. If you’re running multi-agent setups, this category is gold.
Research & Web
Finding and pulling information from the web, APIs, and browsers.
- deep-research — Run deep, source-backed research queries via DeepAPI. Builds a rigorous research prompt, fires it, and saves a cited markdown report.
- deepapi — The Swiss army knife: web search, deep research, scraping (websites, LinkedIn, GitHub, X/Twitter, YouTube), and platform lookups — all through DeepAPI’s endpoints.
- browser-harness — Direct browser control via Chrome DevTools Protocol. Automate, scrape, test, or interact with web pages using the user’s already-running Chrome.
- fireflies-transcript — Pull raw meeting transcripts from Fireflies.ai via its GraphQL API.
- pi-web-search — Web search through Pi.
- online-shopping — Product research and shopping assistance.
Most of these are powered by DeepAPI, which acts as a unified gateway for search, scraping, and research.
Thinking & Docs
Structured thinking, interviewing, teaching, and turning ideas into clear documentation.
These skills handle the “think before you act” problem. Instead of an agent jumping straight to output, these enforce reflection loops, structured analysis, and documentation-first workflows.
Skill Authoring
Creating, improving, distributing, and publishing agent skills themselves.
Meta-skills for building more skills. If the repo is the library, these are the librarian’s tools.
Ops & Setup
Machine, server, security, and tool configuration and operations.
Infrastructure skills. Setting up environments, configuring tools, securing servers. The plumbing that makes everything else work.
What Makes This Interesting
Skills as code, not prompts.
Every skill is a markdown file with frontmatter (name, description, trigger rules) and structured instructions. They’re version-controlled, forkable, and composable. This is a step beyond ad-hoc prompt engineering — it’s prompt engineering as a discipline.
Trigger discipline.
The description field in each SKILL.md doesn’t just explain what the skill does — it defines when the agent should activate it. The cmux skill, for example, explicitly says “Trigger ONLY when the user explicitly says ‘cmux’.” This prevents agent hallucination and skill misuse.
Composability.
Skills can reference each other. The deep-research skill uses deepapi. The orchestration skills reference cmux. You build workflows by stacking skills, not by writing longer prompts.
Hooks directory.
The repo includes a hooks/ directory for agent lifecycle hooks — triggers that fire on events like session start, message received, or task completed. This is where agents become reactive rather than just interactive.
Who Is This For?
- Agent builders who want to stop reinventing workflows and start sharing them
- Power users running agents like Claude Code, Codex, or custom setups who need structured automation
- Teams who want consistent, reproducible agent behavior across projects
How to Use It
git clone https://github.com/davidondrej/skills.gitPick the skills you need, drop them into your agent’s context directory, and reference them from your agent configuration. Each skill’s SKILL.md tells you exactly what it needs and how to trigger it.
The Bigger Picture
The agent skills pattern is part of a broader shift in AI tooling. We’re moving from:
- Chat → Workflow — Agents that don’t just respond, but execute structured processes
- Prompts → Programs — Reusable, versioned, composable instructions instead of throwaway prompts
- Single agent → Orchestration — Skills that coordinate multiple agents, schedules, and tools
David Ondrej’s repo is early-stage but signals where things are going. The 3.3k stars suggest the community agrees.
If you’re building with agents, this repo is worth cloning and studying — not just for the skills themselves, but for the pattern they represent.
Repo: github.com/davidondrej/skills · License: MIT · Stars: 3.3k