Five Skills Worth Installing

Agent skills are easy to misunderstand.

At first they look like plugins.

Install a skill, get a capability. Export a slide deck. Deploy an app. Read a spreadsheet. Generate an image. Talk to a service.

That is useful, but it is not the interesting part.

The interesting part is context control.

The main prompt cannot hold everything

If an agent has dozens of possible capabilities, you cannot put all instructions, examples, tool notes, caveats, and workflows into the system prompt.

You can try.

The result is a huge prompt full of mostly irrelevant instructions. The model has more capability on paper, but less attention in practice.

A skill gives the agent a smaller contract:

Here is the name. Here is the description. Load the details only when needed.

That is a better shape.

Skills are progressive disclosure for agents

Good product interfaces do not show every advanced control at once.

They reveal controls when the user enters the relevant mode.

Agent skills do something similar. The agent sees a lightweight index of capabilities. When the task requires one, it opens the detailed instructions.

This keeps the active context smaller and more relevant.

It also makes capabilities easier to maintain. A skill can carry its own workflow, scripts, templates, and examples without polluting every task.

What makes a skill good

A good skill should not be a vague essay.

It should answer:

• when to use it;
• when not to use it;
• what files or scripts matter;
• what exact workflow to follow;
• what safety checks apply;
• what output shape is expected.

The best skills feel boring in the right way. They turn repeated judgment into a reliable local procedure.

That is why I like the pattern.

It is not about giving the model a costume.

It is about packaging operational knowledge so the agent can load the right amount of it at the right time.