Design systems · AI
Five AI skills I designed that run inside Figma and do the routine work of a design system, from release docs to library audits. One run replaces 0.5 to 32 hours of manual work, depending on the job.
The company
Stellantis is one of the world’s largest automakers, formed in 2021 when FCA and PSA merged. It builds cars for 14 brands. Its UX organization has about 160 designers, and the design system team sits inside it.
Jeep · Ram · Dodge · Chrysler · Peugeot · Citroën · Fiat · Alfa Romeo · Maserati · Opel · Vauxhall · Lancia · DS · Abarth
Try it
Each skill does one job the team used to do by hand. Click one: it runs in about two seconds on a sample library made for this site, and its output appears below.
Example output
LiveReplaces 0.5–1 h of manual work per run, and uses 89–92% fewer AI credits than manual prompting
20–50 credits per run, against 180–600 by manual prompting (proof of concept estimate)
Selected components · 4
Release 2.4.0 / Button / Cover
Loading state added; radius bound to radius.control.
Release
2.4.0 · Oct 17, 2026
Library
Sample DS / Library
Layouts
3 examples
For
Design Review, Release Notes, UX Governance, engineering handoff
How I designed them
01
WhyRelease docs, changelogs, audits, token binding and long engineering specs took hours every sprint.
What I didI listed the routine jobs the design system team did by hand and picked the five with clear rules and the most hours.
SoEach skill does one job, so its output can be checked against a known answer.
02
WhyA model only follows rules it is given.
What I didI wrote each skill as a precise prompt built on the design system’s own token tiers, naming and accessibility rules.
SoOutputs use the system’s own names and limits, which makes them quick to review.
03
WhyA tool designers have to double-check saves no time.
What I didI ran each skill many times against earlier sprint data and tightened the prompt until the output came out the same way every time.
SoFour skills reached a stable output and went live; the UX brief skill is still being tuned.
04
WhyThe skills had to fit how designers actually work.
What I didI interviewed the designers, watched them use the skills and folded their feedback into each one.
SoUncertain changes go back to a designer instead of being applied, which is what made the team trust it.
05
WhyAdoption needs evidence.
What I didIn a proof of concept, I measured manual effort against earlier sprint data and compared AI credits with doing the same job by manual prompting.
SoWith numbers in hand, I shared the skills one audience at a time, from my own team to the global UX team.
8
design system designers use the skills, usually once a sprint
4 + 1
skills live, and one (the UX brief) in progress
16–32 h
of manual work replaced by one sprint audit run; 0.5 to 7 h for the others
60–92%
fewer AI credits than doing the same job by manual prompting, depending on the skill
Hours and credits are proof of concept estimates, measured against earlier sprint data.
How it spread
I built the skills for the design system team. I then shared them one audience at a time, each step reaching more people than the last, until the work reached Figma itself.
01 · 8 designers
I shared the skills with my own DSM team first. Designers started using them, and 8 design system designers use them today.
02 · About 50 people
Next I shared them with the core UX team in North America. Everyone asked me for knowledge sharing, so I ran the sessions.
03 · Product design office
I was invited to the AI team set up across the whole North America product design office, as the ambassador from the UX team.
04 · About 160 people
Then I presented the project to the global UX team. The reactions were very positive.
05 · Next
People now reach out to me for knowledge sharing and training on the project. I will start the trainings soon.
Decisions and lessons
Automation only helps if designers trust what it produces. Most of these calls are about trust: how narrow each skill is, where it works, and when it stops and asks a designer.
Key decisions
Five narrow skills, one job each: release docs, changelogs, sprint audits, token binding and designer briefs.
Skills read components and variables straight from the library and write results back into Figma as frames, reports and bound variables.
Skills apply only the changes they match with confidence and flag the rest for a designer.
Each skill gets the design system’s own rules: token architecture, brand themes and accessibility rules.
What I learned
I ran each skill many times and tuned it until the output was the same every time. A tool designers have to double-check does not save them time.
I interviewed the designers and folded their feedback into each skill, so it fits how they actually work rather than how I assumed they did.
A tool spreads when people see it work on their own kind of work. I shared the skills one audience at a time and ran knowledge sharing sessions when people asked. How it spread is below.