Skills and progressive disclosure

A folder the agent reads when it becomes relevant, instead of a handbook pasted into every request.

Part of the Agents track on lAItest.

You cannot paste the whole company handbook into every request. You also cannot expect the model to guess chapter four.

A skill is a folder

Agent Skills were announced in October 2025 and extended in December 2025 with org-wide management and an open cross-platform format. The format is as plain as it sounds: a folder containing a SKILL.md file that describes a procedure, plus any scripts or reference files it needs. No training, no fine-tuning. The agent reads it.

Progressive disclosure is the whole idea

The agent does not load everything. A short name and description for each skill stay visible, which is enough to know that one exists. The body is read only when the task looks relevant. The bundled files open only if the body points at them. Detail arrives when it is needed and costs nothing the rest of the time.

A common misconception

Commonly believed: Teaching an agent something new means fine-tuning it on your data.

Actually: Most of what people call teaching is procedure, not capability. A file the agent reads at the right moment changes behaviour immediately, can be reviewed in a pull request, and can be deleted when it turns out to be wrong. Fine-tuning is for when the underlying behaviour is wrong, which is rarer than it sounds.

In one sentence

A skill is a document the agent opens when it becomes relevant, which is cheaper than training and far easier to undo.