# Working with an assistant


> The requests that work well, the ones that do not, and how to keep an assistant from doing something you did not mean.
## What works well

**Reading a fleet.** "List every screen in the Frankfurt plant and show which are offline." One call and some filtering.

**Bulk changes through labels.** "Find every screen labeled Warehouse and assign the Safety Rotation playlist to them." Labels are the assistant's lever on a fleet, exactly as they are yours.

**Assembling a playlist.** "Add this video to the start of the Logistics Briefing playlist" is a create then a position update.

**Publishing something it just made.** See [Publish HTML as an app](/docs/developers/mcp/guides/publish-html-as-an-app/).

## What does not

**Anything a tool does not cover.** There is no tool to pair a screen, change a screen's settings, or edit an app's configuration, so an assistant cannot do those however it is asked. See [Tool reference](/docs/developers/mcp/reference/tools/).

**Anything across workspaces.** One token, one workspace.

**Judgement about what should be on a wall.** An assistant will happily put a draft on fifty screens if that is what it was asked for.

## Keeping it safe

**Scope the token** to the access the assistant needs. See [Permissions](/docs/developers/mcp/overview/permissions/).

**Turn on confirmations** if your client supports them. Every destructive tool is annotated, so a client can prompt before deleting a playlist.

**Work on a label, not on everything.** Ask for changes scoped to a label you control rather than to all screens. The blast radius of a misunderstood instruction is then one group rather than the fleet.

**Check Now playing afterwards.** A screen's own page shows what is actually active on it, which is the fastest way to confirm an assistant did what you meant. See [A screen's own page](/docs/screens/overview/the-screen-page/).