Figma's agent went GA. Guidelines are the taste file now

Figma flipped its design agent to generally available yesterday. The headline isn't the badge — it's guidelines.

I've argued that your agent can't see your taste until it's readable — if your rules live in your head, every generation starts from zero. Guidelines fix that inside the library itself. You write the rules once, and the agent reads them on every prompt.

Figma's October 6 agent release makes taste a library setting, not a prompt skill — and solo founders who write sharp guidelines get consistent output while everyone else keeps re-prompting.

What did Figma actually ship on October 6?

A GA agent with memory attached to your system, plus a meter running.

Figma's October 6 release notes marking the agent generally available list the deltas from beta: configurable guidelines per library, search across files from the prompt, live view of teammates' agents on the canvas, and image model choice at the on-canvas entry point. The Verge's report on Figma's design agent going generally available confirms the timing — beta since May, GA now.

Two details matter for solo builders. First, guidelines ship with the library, so anyone who prompts with your library enabled gets your rules in the background. Second, the agent now costs AI credits, which means sloppy context costs real money.

That's a good trade if your files are tight. It's a tax if they're not.

Why do guidelines beat a longer prompt?

Prompts expire. Guidelines persist.

A prompt is a one-session plea: "use my 8px grid, don't invent buttons, keep radius at 10px." It works until you close the tab. Guidelines live on the published library and load automatically, for you and for anyone else with access. You stop paying the prompt tax on every generation.

There's a second win: scope. A prompt tries to cover everything and drifts. A guidelines file covers one library and stays checkable. The agent doesn't guess your intent from adjectives — it reads "use card/product with 24/40 padding, never card/legacy" and follows it.

I don't think this kills prompting. It moves prompting up a level. You prompt for the job — "draft checkout empty state" — and the guidelines handle the how.

What goes in a guidelines file that actually works?

Short files about decisions the agent can't infer from pixels.

Figma's guide to adding guidelines files to a design library recommends three to eight focused files under 100 KB combined — overview.md, components.md, anti-patterns.md, that shape. That limit is a feature. If you need more than eight files, your system is the problem.

Here's the format I use:

• Situation. When does this rule apply? "Product cards in browse grids."
• Do this. Name the exact component and tokens. "Use card/product, zinc-900 on white, 10px radius."
• Not that. Name the tempting mistake. "Don't use card/legacy or add drop shadows above 3%."
• Why. One sentence. "Legacy cards break grid rhythm on mobile and fail contrast in dark mode."

Write from your review history. Which corrections do you repeat? Confused primary versus promo buttons, invented spacing, off-brand empty states — those are your first three files. Ask the agent to draft them from your docs, then cut hard. Keep component specifics in asset descriptions and cross-library judgment in guidelines. Don't repeat the same rule in both places; duplicates burn credits and confuse routing.

How do you keep the agent honest after it ships?

You review output against two gates, then update the file.

System gate first. Did it use only your components and variables? If it invented a new button, kill it or promote it deliberately. Silent variants are how drift starts.

Judgment gate second. Does the screen have one job, clear hierarchy, and copy that sounds human? Token compliance isn't taste. A correct grid with vague copy still ships doubt.

My cadence: one watched generation per week. Same task, same library, fresh prompt. Screenshot the misses, add one rule to the guidelines file that day, re-run. Teammates' live agents help here too — when you see someone else's agent invent a pattern, that's a missing rule, not a bad user.

Figma gave the agent your library. Your job is to make the library opinionated enough to deserve it.