The Figma agent produces work that reflects your product only to the extent that it knows how your team builds. Most of that knowledge already lives in your design system: the components and tokens it should reach for, plus the documented rules and examples showing how they combine. The task is making that context reachable when the agent needs it.

Connect the library first

Component and variable libraries carry the raw building blocks. Once a published library is connected, the agent can tell which assets your team has approved and draw on those components and tokens instead of inventing its own styling and structure.

To connect one, open an agent chat, select “Libraries” in the prompt box, and pick the library the project should follow. In a checkout flow, an order summary card request might resolve to the existing Card/OrderSummary/Default component, the approved button variant, and your spacing and color variables. When the change is narrow — swapping a single button on a screen, for instance — @-mention the component or token you want the agent to apply.

The quality of the signal depends on the library itself. Meaningful component names, defined properties and variants, applied variables, and reusable patterns that show how smaller pieces fit together all help the agent understand what an asset is for and how it should behave.

Codify judgment with guidelines

The library states what exists, not when to use it. Guidelines are written instruction files attached to a published library, and they are pulled in automatically for any prompt where a library is enabled, so nobody has to restate the rules each session.

Useful content for a guidelines file includes:

  • Rules the agent should follow consistently
  • Exceptions to those rules
  • Mistakes to avoid and the correct alternative
  • How to choose among similar components, variants, or patterns

The attachment procedure runs through the design library: Library → Manage guidelines, upload the file, then click Apply changes. A typical rule might reserve the primary button for the “Place order” CTA on checkout while secondary buttons handle actions like “Back.” Because guidelines stay with the library, the guidance is reusable across new work. They perform best when the agent is building from scratch; Figma says it is working on making their application more consistent when the agent edits existing designs.

Four kinds of context, one mental model

These are the reference types available to the agent:

  • Guidelines — standing rules for how the team uses the library; attached to it and applied to new work.
  • Library examples — finished designs showing how components and patterns fit together in practice.
  • Add context — references tied to the current task, such as a particular Figma file or frame.
  • Skills — a repeatable process the agent should follow.

An example often carries more weight than a prompt-written description, much as a colleague would rather start from an existing checkout screen than from an explanation of one. To hand the agent a design, click + → Attach Figma files and select or paste it. Patterns worth reusing should be published as complete components on an Examples page, or have _example appended to their names — both are case-sensitive, and for component sets the naming or page placement belongs on the parent set rather than an individual variant. The Material 3 Design Kit is a reference implementation. One-off references that won’t recur can simply be attached to the current prompt, pasted into the chat as a frame or image, or turned into an editable design from a screenshot.

Consider a shipping address form being added to the checkout flow. The library supplies the fields and checkboxes, but says nothing about which fields share a row or how optional entries like apartment numbers appear. A previous design demonstrates that grouping and spacing, giving the agent something to adapt.

Skills for repeatable processes

Skills are instruction sets taught once and triggered in any chat with a forward slash (/). Create one by describing the desired task to the agent along with the steps to follow; the result can be previewed and edited, including directly in the markdown editor. Skills can also be pulled from the Figma Community or shared with others.

Audits are a natural fit. A /design-system-audit skill could check selected screens against your libraries and flag detached instances, hardcoded values, misused components, and accessibility gaps. An /analyze-components skill could look for duplicate components, naming errors, and components no longer connected to the library. Both surface problems with a recommended path, which the team can have the agent execute or handle directly.

Correct on canvas, then fix the input

Mistakes can be fixed in place: describe the change in a prompt or @-mention the exact component or token. As with any AI output, review still matters. A correction that keeps recurring is a signal to revise the guidelines, revisit the library, or supply a better example rather than to keep patching outputs.

Seat availability: the agent is available to Full seat users on Professional, Organization, Enterprise, and some Education plans, with limited availability on Starter. Collab, Dev, and View seats can use the agent in their drafts.