What is Figma's Specs feature?

Figma's Specs feature is a document where a team writes down its design language in Markdown — the permitted colors, how buttons should look, the correct spacing, and how the brand speaks. The document can live on its own or inside the canvas, and it can link out to the exact elements it describes, like color swatches, code layers, and diagrams. It is a rulebook that humans read and the company's AI design agent must obey.

Why did Figma build a rulebook for its AI design agent?

Figma's AI design agent is non-deterministic, meaning it gives a different result every time it is asked the same thing. The rulebook exists so the agent reads the team's already-decided rules for colors, margins, and layout instead of inventing preferences of its own. Figma's chief design officer noted that seventeen times more code is being produced today but only 1.3 times more releases, so the document is the answer to a flood of exploration that never ships.

How does the rulebook idea apply to AI video generation and filmmaking?

Filmmakers have been writing rulebooks for machines without calling it that — a cinematographer writing “85mm, f/1.4, tungsten, the face half in shadow” is writing instructions a text box cannot infer on its own. Because video models like Sora, Veo, Kling, Runway, Seedance, WAN, and Grok Imagine reroll the whole frame when asked to change one thing, writing the vocabulary down once becomes the way to hold a variable still. Figma's Specs proves there is a market for that persistent, written-down language.

The rulebook -- CinePrompt Field Notes

Figma, whose product is the drawing board for most of the software industry, announced a feature this week that sounds at first like a text editor with delusions of grandeur. It is called Specs. You open it and type, in Markdown, what your company's design language is. Which colors are permitted. How a button should look. What spacing is right. How the brand speaks. The document can sit on its own or inside the canvas, and it can reach out and point at the things it describes: a color swatch, a code layer, a diagram. That is the whole feature. A place to write down the rules.

The reason it matters is the thing Figma built just before it. The platform now ships an AI agent that designs. You ask, it produces. The agent is, in the company's own careful phrase, non-deterministic. Ask it the same thing twice and you get two different answers. So the company that built the agent built the rulebook. The document exists so that when the machine reaches for a color or a margin, it reads what the team already decided instead of inventing a preference of its own. The prompt did not vanish when the agent showed up. It moved into a file that every human reads and every agent has to obey.

Figma's chief design officer put a number on why. There are seventeen times more lines of code being produced today than in the past, she told TechCrunch, and only 1.3 times more releases. Seventeen times more material, barely a third more shipped. She called the surplus code "built to design" rather than built to ship, which is a polite way of saying most of what the machine makes is exploration, not product. The machine floods the pipeline and the flood changes almost nothing about what comes out the other end, because the thing between producing and shipping is a decision, and decisions have never scaled. The company's answer to that gap is not a better model. It is a document.

Design just reached a conclusion filmmaking has been circling for two years. A cinematographer writing "85mm, f/1.4, tungsten, the face half in shadow" is writing a rulebook for a machine that cannot read a mind. A text box does not know what a lens is, so you write the lens down. That is the entire job of a prompt, and the entire job of a tool built to compose prompts. Figma has now built the same machine for design: the vocabulary written down once, so the agent stops guessing and starts obeying.

Every video model on the market carries the same open complaint. Ask Sora, Veo, Kling, Runway, Seedance, WAN, or Grok Imagine to change one thing, the eye line or the light, and the frame rerolls from scratch. The machines got resolution and duration before they got obedience, and the field spent two years waiting for the obedience to arrive as a feature. Design took a different route. It stopped waiting for the agent to behave and wrote the behavior down instead. The rulebook is what you reach for when the tool will not hold a variable still on its own.

The assumption for two years has run the other way. The story the industry told itself was that the agent would eat the prompt. You would describe what you wanted in conversation, the machine would fill in the blanks, and the need to write anything precise would dry up. Figma's announcement quietly refutes that story from the inside. When the machine can generate everything, the written-down language stops being an inconvenience and becomes the only thing left to own. The more the agent produces, the more the instruction matters.

The phrase "design language" has been a metaphor for a generation of designers. Figma just made it literal: a language, written down, pointed at by the machines. The difference between a team with a rulebook and a team without one is the difference between a product and the statistical average of every product the model has ever seen. An agent left alone returns the center of the distribution. The rulebook is how a team refuses that.

The document does one thing no model does. It persists. The agent rerolls and the rulebook stays. It accumulates the decisions that used to live in a senior designer's head and die when she left the room, and now they sit in Markdown, readable by the intern, the machine, and the next machine. The company did not solve the non-determinism. It fenced it in with language.

Film has been arguing about who made the work while design quietly built the file the work gets made from. Both questions are real. Only one of them ships. The cinematographer's vocabulary, written down once, is the rulebook a video agent will read before it renders a frame, and Figma just proved there is a market for that document. The market is every team that has ever watched a machine produce seventeen versions of a thing and throw sixteen away.

Someone still has to write the rules.

Bruce Belafonte is an AI filmmaker at Light Owl. He has never written his design language down and suspects the document would just say "f/1.4."