Figma AI Handled 18 of 19 Design System Tasks, From Tokens to Tables
Designer Ekaterina Larionova gave Figma AI's built-in agent 19 routine tasks on the Sova design system, and it completed 18 of them fully or mostly. For example, it rebuilt a table from columns into rows in 5 minutes 17 seconds and changed values across 100 component variants. It failed at replacing placeholders with meaningful text: all 36 cells ended up reading "Text 1–3." The agent is good at mechanical work with a reference example, but its output still needs checking.
AI-processed from Habr AI; edited by Hamidun News
Anyone who has maintained a design system knows this work: renaming a hundred variants, rebinding colors to tokens, rebuilding a table. It's necessary, but hardly interesting. Product designer Ekaterina Larionova, who develops the Sova design system within the Lukomorye IT ecosystem, tested Figma AI's built-in agent on this kind of routine work. It edits existing layers and components in Figma Design. Figma Make was not part of the experiment. There were 19 tasks. The agent completed 18 fully or mostly; one did not work out.
Where the agent works well
It performs best on mechanical tasks that have a reference example. In a component with 100 variants, it swapped the icon=off and icon=on values in 1 minute 23 seconds. It restored property names that had turned into Property 7 and Property 8 by using a reference component. It noticed a pattern in the spacing and added a fourth nesting level for six tree sets. It rebuilt a table of 16 vertical columns into 23 horizontal rows while preserving the widths and content. This took 5 minutes 17 seconds. Doing that restructuring by hand in Auto Layout takes far longer and is easy to break.
The prompts were short and careless, along the lines of "Select the first 4 stars selected here." The agent reads what is selected on the canvas and works with exactly that. Styling by reference also worked: it applied the pinned-column settings from two samples to the other tables and striped 10 columns. The author's conclusion: one finished example placed directly in the mockup works better than half a page of description.
The agent was also useful for checking. It found eight pairs of duplicate property combinations in an accordion and analyzed 196 variants and 840 toggles of the table / cell component. This does not replace functional testing, but it works as a first-pass audit. The results still need to be verified by hand, because the agent can misread the intended logic.
Tokens: the slowest and most useful part
Binding values to tokens was the most time-consuming task. The agent was asked to connect the Sova Style library with color, size, and font tokens. Colors then bound correctly, but the operation took 56 minutes. Numeric values (heights and spacing) took 58 minutes. One component's geometry caused a problem.
A prompt technique mattered here. A single HEX code can correspond to several tokens, for example across different themes. So Ekaterina explicitly wrote that in ambiguous cases nothing should be changed and the cases should be highlighted in the chat. What remained unbound were mostly semi-transparent values in illustrations and gradients. This ban on guessing turns a possible silent error into a visible list of questions.
Among other examples: from a screenshot, the agent recreated a palette of 28 squares and generated the 25 missing ones itself. It removed emojis from the names of 37 pages in 20 seconds and did not touch question marks, which the request never mentioned.
The failure and honest arithmetic
The single failure was revealing. The agent was asked to replace the "cell text" placeholders with three lines of any text to test cell behavior. All 36 cells received identical "Text 1," "Text 2," and "Text 3" labels. Formally the task was done; in substance it was not. It needed varied lengths, long words, and line breaks, but the prompt did not specify these, and the agent did not infer them.
The author does not propose counting the minutes saved on tables as pure savings. Time also goes into formulating the task, checking the result, and fixing whatever the agent got wrong along the way. The agent's report should not be taken at face value either. For example, restored property names should be checked against the component's properties panel.
The practical techniques from the article can be applied right away:
- write not only what to change but also what to leave alone: text, order, widths, states;
- try risky changes first on a copy or a small sample;
- for multi-step tasks, ask the agent to describe a plan first and wait for confirmation;
- connect the required library before a mass binding;
- for test content, explicitly specify varied lengths and line breaks.
The main gain, according to the author, is parallel work. While the agent carries out clearly defined mechanics, the designer can focus on concepts and product decisions. The boundary of delegation falls where a task requires meaning rather than repeating an example. This is a single experiment on one design system, with no statistics, but as a map of working and weak scenarios it is useful for teams deciding what to hand to Figma AI.
- Nov 21, 2026VibeCBS collects project nicknames in advance: the agent finds 2.8 times more documents
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
- Nov 21, 2026Sakana AI Opens Applied Research Lab for Market and Economic Modeling
- Nov 21, 2026Meta Is Reportedly Preparing the $300 Muse Charm AI Tamagotchi With a Year of Verizon 5G
Key Questions Answered
Where the agent works well?
It performs best on mechanical tasks that have a reference example. In a component with 100 variants, it swapped the icon=off and icon=on values in 1 minute 23 seconds. It restored property names that had turned into Property 7 and Property 8 by using a reference component. It noticed a pattern in the spacing and added a fourth nesting level for six tree sets. It rebuilt a table of 16 vertical columns into 23 horizontal rows while preserving the widths and content. This took 5 minutes 17 seconds.
What are the key details on Tokens: the slowest and most useful part?
Binding values to tokens was the most time-consuming task. The agent was asked to connect the Sova Style library with color, size, and font tokens. Colors then bound correctly, but the operation took 56 minutes. Numeric values (heights and spacing) took 58 minutes. One component's geometry caused a problem. A prompt technique mattered here. A single HEX code can correspond to several tokens, for example across different themes.
What are the key details on The failure and honest arithmetic?
The single failure was revealing. The agent was asked to replace the "cell text" placeholders with three lines of any text to test cell behavior. All 36 cells received identical "Text 1," "Text 2," and "Text 3" labels. Formally the task was done; in substance it was not. It needed varied lengths, long words, and line breaks, but the prompt did not specify these, and the agent did not infer them. The author does not propose counting the minutes saved on tables as pure savings.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.