# Skills and plugins (/skills-and-plugins)


A **skill** is a package of instructions for a kind of work — how to build a
spreadsheet that opens cleanly in Excel, how to validate a chart before
claiming it is done, which library to install and at which version. A
**plugin** bundles related skills under one name.

Skills are what make the agent's document work reliable. Without them, a model
asked for a PowerPoint improvises.

## What ships with the app

| Plugin | Skills |
| --- | --- |
| **Documents** | Word documents, PDF documents, presentations |
| **Spreadsheets** | Spreadsheets |
| **Charts** | Charts |

Built-in plugins and skills you write yourself are enabled by default. A
plugin installed from a public archive starts off — turn it on in the Plugins
panel before the agent can use its skills or launch any bundled MCP server.
The agent sees a catalog of the enabled skills and reads a skill's full
instructions when it decides to use one.

## Using one explicitly

Type `/` in the composer to open the plugin library and pick a skill. It
attaches to the message as a chip, telling the agent to use that skill for
this request. The **Plugins** entry in the sidebar opens the same library as a
full panel, with search.

You do not have to do this — the agent picks skills on its own. Invoking one
explicitly is for when it picked wrong, or you want a specific output format.

## Enabling and disabling

Disabling is install-wide, not per chat. A disabled skill is not staged into
the sandbox and not advertised to the model, so it is genuinely gone rather
than discouraged. Disabling a plugin disables its member skills without
forgetting your per-skill choices — re-enable the plugin and they come back as
they were.

## Prompts

Saved prompts live alongside skills in the same library and are also reachable
with `/`. A prompt is text inserted into your message. It is never added to
the agent's operating instructions and never staged into the sandbox — it is
a shortcut for you, not an instruction to the model.

## Writing your own

Skills, plugins, and prompts are Markdown files with YAML frontmatter, read
from your application data directory:

```text
<data dir>/skills/<name>/SKILL.md
<data dir>/plugins/<name>/PLUGIN.md
<data dir>/prompts/<name>/PROMPT.md
```

A `SKILL.md` needs a `name` and a `description`, and may declare dependencies
the sandbox should install:

```md
---
name: charts
description: Produce charts as interactive native figures, or as PNG/SVG when an image file is needed.
deps: { python: ["matplotlib==3.9.4"] }
---

# Charts

Instructions the agent reads when it uses this skill…
```

`deps` takes `python`, `npm`, and `host` lists. Pin versions.

A `PLUGIN.md` names the skills it groups:

```md
---
name: documents
display-name: Documents
description: Word, PDF, and PowerPoint deliverables, each validated before delivery.
category: documents
skills: ["word-documents", "pdf-documents", "presentations"]
---
```

Your skills are loaded after the built-in ones and cannot shadow a built-in
name. Pick a different name if it collides.

<Callout>
The description is what the model sees in the catalog when it is deciding
whether to use your skill. Write it as selection criteria, not as a summary.
</Callout>

## Capability badges

The app labels each package with what it actually does — whether it writes to
the workspace, needs network access, requires a host program, or brings an MCP
server. Those labels are derived by reading the package. A package cannot
declare its own badges, and a manifest that tries to is rejected.