Yana Tarnovetska
← All work

Designing with AI

I built a set of reusable AI skills inside Figma, each with one job and clear rules, so AI speeds up the work and I keep the decisions.

Figma agents

Context

As the only designer on 10+ products, I kept facing the same trade-off: move fast, or do it right. Skipping problem framing led to designs built on unclear goals. One draft per review led to several rounds with PMs. Without peer review, usability issues reached analysts. I built AI skills to cover these steps, so I didn’t have to choose between speed and quality.

Who it affects
My PMs, PO and engineers across 10 products, and the analysts who use those products

Actions

How I got there

  1. Split the work into skills with one job each

    Instead of one big prompt, I built a set of reusable skills with an agent. Each skill has one job and clear rules, grouped into three types.

  2. Understanding: turn a messy problem into goals

    Before I design, I run skills like context scenario and job stories. They help me turn a messy problem into user goals and requirements.

  3. Exploring: 2 to 5 options, with trade-offs

    When there's no time for full wireframes, a skill gives me 2 to 5 layout options. For each one it lists the positives, the negatives and the effect on users. Then I pick a direction and explain why.

  4. Critique: feedback only, never edits

    Critique skills check whether a design is easy to use, follows proven UX patterns, and presents data clearly.

    Why no edits: if a critique skill could change the design, I'd be reviewing AI decisions instead of making my own. Feedback-only keeps me in control.

Results

12
skills that automate structured design tasks
10–40 min
of manual work saved each time a skill runs
4–5 h
reclaimed per week

For my team

The same structure every time

PMs, PO and engineers

Every review, handoff and critique follows the same structure, so quality doesn't vary from one session to the next.

More options in the same time

PMs and PO

Design options explored per cycle went from 1–2 to 3–5, in the same time budget.

Faster design system docs

Engineers

Documenting a component went from 10 minutes to under 1 minute.

Learnings

I thought AI would make me faster. The bigger change was consistency. As the only designer, my reviews were worse in busy weeks. Now every review follows the same steps, so PMs know what to expect.

I used to bring one design to a review and defend it. Now I bring several, each with its pros and cons. I learned that more options only help if I can explain which one fits best and why. AI gives me more ideas. Choosing is still my job.

Most importantly, I learned that AI sounds just as sure when it's wrong. Now I check its most confident answers first. I build the same idea into AI tools for analysts: make the AI easy to check.

Building this site in code showed me the same thing. The code always looked finished, but when I clicked through it:

  • a case study showed a result I had never measured
  • the page loader ran only on the home page
  • the slider divider disappeared against the screenshots
  • spacing didn't show what belonged together
  • in the dark version, headings and body text looked the same
  • screenshots were cropped on my laptop

I only caught these because I reviewed every page after each change and asked where every number came from. Now I check every claim against a real source, and every screen against the design system, before I trust them.