A skill packages instructions, references, assets, and optional scripts so an agent can reliably perform a specialized workflow when the right task appears.
A technician grabs the kit labeled “fiber repair” only when a fiber line breaks. Inside are the procedure, diagrams, and specialized tools.
An agent sees a skill’s name and description, then opens its full instructions and supporting files when the task matches.
The skill is not the model. It is a reusable package of context and procedure the host can give the model.
This is progressive disclosure: keep the catalog lightweight, then spend context only on the capability being used.
The presentation workflow is relevant, so its full instructions become available.
A strong skill description states the job, trigger phrases, scope, and boundaries. The body explains how to do the work.
Use for a single request or lightweight preference that does not need a maintained workflow.
Use when the agent needs repeatable judgment, ordered steps, templates, or domain guidance.
Use when the agent must query a service, mutate external state, or perform a deterministic operation.
A mature workflow often combines them: the prompt supplies intent, the skill supplies method, and tools supply capabilities.
Overly broad triggers load the wrong workflow; vague ones may never activate.
The skill must respect higher-priority user, system, and safety requirements.
Version domain guidance and verify unstable facts against authoritative sources.
Validate inputs, preserve user data, and keep side effects explicit.
The lightweight label lets the agent decide whether the skill fits.
The full workflow appears only after the skill is selected.
Supporting material enters the process only when the workflow calls for it.