On November 20, Figma published a workflow walkthrough in the Workflow lab section of its official blog, written by Johnny Dallas, the technical lead for Specs. Using a fictional team, the post shows how an engineer can use Specs and the Figma agent to gather scattered requirements, data and designs onto one canvas, produce an implementation plan a coding agent can follow, and then hand it to OpenAI's Codex to write the code. A note at the end says Specs is in public beta, available on all plans, on web, desktop and mobile.
What Specs is
Specs is a new canvas element from Figma, described officially as a "container for context." Its content is written in Markdown and sits on the same canvas as the designs it describes, and both teammates and agents can write into it. Alongside it comes code layers: interactive prototypes on the canvas driven by real code, also in public beta.
Figma's reasoning is that teams now move from prototype to launch faster and faster, yet written requirements, design details and engineering conventions are spread across different documents and links, and are easily lost at handoff.
The five steps in the demo
The example company is Pickup, a fictional shopping-platform app that is building a delivery feature.
- Fill in the data first. The requirements doc says to pilot in one city but does not say which. The engineer connects the Figma agent to the Hex analytics platform through a connector and has it find the cities with the highest usage. It writes the result up as a spec with a table and generates map code layers for the top four cities. After reviewing them, the team picks Bengaluru, India.
- Gather the materials. The technical proposal, the delivery flow chart in FigJam, the annotated final designs, the interactive Bengaluru map, and the new components from the design system team along with their usage notes all go onto the canvas.
- Generate the plan. The engineer selects these items and runs a build-implementation-plan skill the team had set up in advance. The agent produces a new spec describing how the feature looks and behaves, the testing and monitoring requirements, the rollout schedule and a task list, and marks which source material each point comes from.
- Human review. The engineer leaves a comment on the plan pointing out an error and asks the agent to read it and reply with a fix. Another point needs a decision from the designer, so the engineer @-mentions them in a comment and the designer steps in and edits it directly.
- Hand off to Codex. The engineer copies the spec's link into Codex with a single line: "Implement this feature." Through Figma's MCP server, Codex reads not only the plan itself but also the annotated designs, the code layers and the components.
Throughout the flow, the Figma agent writes no production code. It fetches data, organizes material and drafts the plan, leaving the coding to an external coding agent.
What the demo leaves unanswered
This is an official demo, and Pickup is an invented company, so the post contains no numbers from any real team. It does not say how long plan generation takes, how much rework Codex needs after following the plan, or how the agent handles materials that contradict one another. It also does not mention how Specs will be priced once the public beta ends.
Is it useful for teams in China?
For teams in China, this workflow faces two hurdles.
The first is upstream. The agent and connectors depend on overseas models and overseas services such as Hex, while tools Chinese teams commonly use, such as Feishu docs and Yuque, are outside the demo. Whether they can be connected depends on whether the MCP ecosystem has a matching service.
The second is downstream. The demo uses Codex, but developers in China more often work with Tongyi Lingma, TRAE, or open-source coding tools wired to domestic models. MCP is an open protocol and, per Figma, only a link is needed, so in theory other MCP-capable coding tools could read the same context. But the post demonstrates only Codex, and Figma offers no official word on how other combinations perform.
Domestic design tools such as MasterGo have also offered MCP interfaces for some time, handing design structure to coding tools to read. Figma goes one step further this time by bringing requirements documents, data conclusions and design system notes into the same entry point.
Sources: Figma official blog, Workflow lab section, used to verify the demo flow, public beta scope and supported platforms; CocoLoop; Figma product documentation on Specs and code layers, used to verify the definitions of the two features.