the ai features in figma, what they actually do, and when they save time vs waste it

design · 2024-11-29 · 7 min

figma added a lot of ai in 2024. some of it is genuinely useful. some of it is a demo that costs more time than it saves. i used the features for a month on real files, including one file with four years of layer debt, and kept notes.

this is the split, task by task. no benchmark, no screenshots of a perfect result. just what happened when i tried to get work done.

rename layers

the feature: select a frame or a messy group, ask figma to name the layers.

the verdict: useful, with a caveat.

on a file where every layer is called "frame 427", it turns a fifteen-minute cleanup into a two-minute review. that is a real win. on a file with a naming convention, it invents names that are almost right, and "almost right" is worse than wrong because you have to read every one to catch the difference.

the caveat: it names what it sees, not what the component means. a rectangle with a border becomes "card border", even if it is actually a focus ring. the feature is a first pass, not a source of truth. use it on inherited files, not on your own system.

generate text

the feature: ask for copy inside a design, usually a headline or a paragraph.

the verdict: useful for placeholders, bad for final copy.

for a wireframe or a stakeholder review, generated text is better than lorem ipsum because it gives the layout realistic length and rhythm. for anything a user will read, it is a draft written by something that has never met the user. the tone is generic, the claims are vague, and the specificity that makes copy good is exactly what the model cannot invent.

i use it to fill a card so i can see whether the card works. i replace every word before the file leaves my machine.

remove background

the feature: cut a subject out of an image.

the verdict: the most consistently useful feature in the set.

it is fast, it is good enough for a mockup, and it fails in predictable places: hair, transparent objects, soft shadows, anything that touches the edge of the frame. for a product shot on a clean background it saves a trip to another tool. for a portrait with flyaway hair it saves nothing, because you spend the saved time fixing the mask.

the honest use is the mockup, not the final asset. for the final asset, use a proper mask and check the edges at 200%.

auto-layout suggestions

the feature: figma proposes auto-layout, spacing, and structure for a selection.

the verdict: occasionally magic, often noise.

when it works, it turns a stack of rectangles into a responsive card in one click. when it does not, it produces a structure that is technically valid and semantically wrong: a button made of three nested frames, a list that is actually a grid, padding applied to the wrong container.

the pattern i noticed: it is best on simple, common patterns and worst on the thing you are actually designing. that is not a coincidence. the model has seen a million cards. it has not seen your card.

first draft

the feature: describe a screen, get a layout.

the verdict: the biggest time sink of the set.

for someone who does not know figma, a first draft can be a starting point. for someone who does, the draft is a puzzle made of someone else's decisions. you spend the first ten minutes deleting, the next ten renaming, and the last ten rebuilding the thing you already knew how to build. the demo is impressive because the demo has no existing system to fit into. real work always does.

asset search

the feature: search the file or the team library in natural language.

the verdict: quietly excellent.

"the blue icon with the arrow" finds the blue icon with the arrow. "the card from the pricing page" finds the card from the pricing page. this is the least glamorous feature and the one i use most, because it removes a navigation task rather than making a design decision. it also gets better as the library gets cleaner, which is a nice incentive.

translate

the feature: translate text layers into another language.

the verdict: a good first pass, a bad final pass.

for a quick review of whether a german headline fits, it is useful. for shipping, it produces text that is grammatically fine and culturally generic. the words are translated; the intent is not. if the copy matters, a human translator is still cheaper than a bad launch.

the rule i kept

use ai for the boring middle. not the beginning, where the decision lives, and not the end, where the judgment lives. renaming layers, cutting out a product shot, finding an asset: boring middle. choosing the layout, writing the claim, deciding what the screen is for: not the middle.

figma's ai is good at the first category and confident about the second. the confidence is the problem, not the capability.

for the variables side of the same system, see figma variables explained. for the drift between design and code, see figma component drift.