Mental Model

A skill is a capability package the agentAgentA worker you delegate to: brief it, and it takes steps on its own. A chatbot answers; an agent acts.An LLM that runs tools in a loop toward a goal — it acts, checks the result, decides the next step, and repeats until done.Full definitionClose can discover and load when a task needs specialized knowledge or procedure. It moves long, conditional guidance out of the always-present promptPromptThe brief. Closer to writing instructions for a new contractor than to programming.The text you give the model — question, instructions, context. Its quality shapes the quality of the output.Full definitionClose.

Three Layers

1. Registry metadata — name and description used to decide relevance.
2. Instruction body — the workflow, decision points, and completion contract loaded on invocation.
3. Resources — scripts, references, templates, and assets loaded or executed only when required.

This progressive-disclosure shape is portable across Claude Code, Codex, and Copilot. Discovery paths, precedence, arguments, allowed-tool syntax, model controls, and compaction behavior are not.

Activation Is a Routing Decision

Model-invoked skills trade operator effort for probabilistic routing. User-invoked skills provide timing but still depend on the model to execute the procedure. A skill description should discriminate when the workflow is required; the body should state observable completion criteria.

Capability, Not Authority

A skill can teach, scaffold, inspect, or orchestrate. It should not grant itself new authority. External writes, destructive changes, publication, and spend remain subject to the surrounding product permissions and human decision boundary.

The crux is simple: if skipping a step would make the result invalid, verify the step's output independently instead of assuming the model followed the skill.