Kelly Mears

Agent Skill

A packaged, invocable set of instructions that loads into an agent's context on demand.

Agents & Language Models2 min read223 words9 out · 4 in
also calledSkillSlash command

An agent skill is a reusable bundle of instructions — often a single document plus supporting reference files — that an assistant loads when a task matches it. It is the unit of packaging for procedural knowledge: how to run a particular workflow, what conventions apply to a kind of work, which steps a task requires and in what order.

The design pressure is the same one that governs all context: a skill that is always loaded is just a longer System Prompt and competes with everything else for attention. Loading on demand — triggered by name or by a description matched against the task — is what makes a large library of them affordable.

Skills tend to accumulate the procedural material that would otherwise bloat a project's standing instructions. Keeping the standing instructions lean and moving step-lists into skills is a recurring refactor, and it has a corollary: standing instructions should say what is true, skills should say what to do.

Practical hazards are those of any plugin system. Loading is by convention — a bundle placed where the loader does not scan is silently inert, which is a Silent Failure with no error anywhere. Names must match directories exactly. And a skill that names a specific model or a specific file layout goes stale in the ordinary way; see Documentation Rot.

See also4

Hand-picked in the note itself — the neighbours worth reading next.

Related3

Nearby in the graph rather than deliberately chosen. Looser, sometimes surprising.

Linked from4

Notes elsewhere in the wiki that reach for this one.