Skills

Agent skills are the most powerful and versatile harness primitive. At their most basic, a skill is a SKILL.md file of instructions for handling a specific task or workflow, loaded when the model identifies that the user’s request calls for it. Like AGENTS.md, skills exist at both user (global) and repo level.

What skills are for

Two broad shapes have emerged:

• Task/workflow skills - step-by-step instructions for niche or personal processes: a deploy ritual, a review checklist, a content pipeline.
• Domain skills - context for a concept area: frontend design, creative writing, code review, multi-agent orchestration. These fire automatically when the request touches the domain.

The same discipline as AGENTS.md applies: skills should predominantly address observed failure modes. Models still struggle to produce high-quality novel designs, so a frontend-design skill that injects design context earns its keep. A skill restating what the model already does well is dead weight.

Writing skills that get followed

Compliance varies - some models follow a skill to the letter, others treat it as advisory. To maximise your odds:

• Keep language tight and explicit; avoid vagueness that forces the model to infer intent.
• Prefer concrete rules (“never start article content with an H1”) over aspirations (“write engaging content”).

Repo-level skills

In a codebase, skills help a model work seamlessly without manual context-gathering: navigating to the right modules, team working patterns, code and testing standards, and how to operate infra providers, CI, and related repos. They’re the difference between an agent that works in your repo and one that works on it.

Start from the best public collections - mattpocock/skills for engineers, emilkowalski/skills for designers, coreyhaines31/marketingskills for marketing and sales - and adapt rather than write from scratch.