AI prototyping kit
In 2026, adopting AI became a KPI for the design department, but off-the-shelf tools didn’t know our components or what the team had agreed on. I built a lightweight design system for vibe coding: a React kit with rules for the agent, a canvas for edits and a knowledge base. Now a prototype on Banki.ru components starts from one file, and edits go to the agent as comments.
About
- Role
- Product designer, author of the kit, canvas and knowledge base
- Period
- March–November 2026
- Status
- Rolling out to the design team
In numbers
- 1 file
- for the agent to spin up a new prototype
- 46
- component guides written for the agent
- 30–40%
- fewer tokens on prototype edits
Problem
The agent didn’t know our components or what the team had agreed on for a task. Prototypes ignored the design system and lacked context, so they were redone by hand. Every redo cost time and tokens. The task was for the agent to build prototypes on our components and with the task’s context from the start.
What we tried
Figma Make didn’t understand the design system, and Figma agents followed it poorly even with guides. Cursor with Figma MCP built a page in 5 minutes but burned lots of tokens, and the quality was questionable. A multi-screen prototype through OpenCode and the API cost about $150.
How it works
Three parts, each answering its own question for the agent. The kit: how to build a screen. The knowledge base: what we’re doing and why. The canvas: what to fix. The designer works in Claude Code, and the agent takes all three from one folder.
A prototype from one file
A new prototype starts with NEW-PROTO.md. The agent asks only for a name, installs dependencies, sets up the project with git and the kit, and runs it. It replies with a link and instructions for deploying to Vercel.
A kit with rules for the agent
The kit has 31 components, 20 organisms and 8 page templates in React. They mirror the production Storybook: developers see the same component names in a prototype as in Figma’s inspector. Each has a guide for the agent — when, why and how to use it, and why something is off-limits. The agent may not change component parameters unless the designer asks.
States without separate mock-ups
Organisms have a state configurator: for example, signing with a button or a code widget, for a microloan, loan or deposit. So every screen state comes from one prompt instead of a mock-up per state. In Figma that would be a dozen frames to maintain by hand.
A canvas with edits as comments
Coded screens sit on a board like in Figma: live, in phone and desktop frames, with teammates’ cursors. The designer presses C, drops comments on any screens and copies them all into the chat with the agent. Each comment automatically carries:
- the element and the designer’s note
- the frame and screen it sits on
- the styles currently applied to it
- the file and line where it lives in the code
Knowledge base
The kit describes how to do things; the knowledge base covers what and why. It holds about 320 research studies since 2019, task cards built from meeting transcripts, analytics, metrics and copy rules. The agent sees what the team has already decided and what worked in testing, and doesn’t wipe it out in a prototype.
Research from PDF to text
Since 2019, every study was a PDF slide deck stored in Confluence. Agents can’t really read those: numbers hide in charts, text sits in images. So I built an agent pipeline that turned the whole archive into markdown:
- each deck is cut into slide images
- one agent per slide transcribes text, charts and tables
- other agents check a random sample of slides
- a second pass rewrites the text into one structure with headings and checks the numbers again
Search by meaning
The whole archive took 30 minutes and about $5. Each document is cut into chunks by headings, and an embedding model turns them into vectors. They live in Postgres on Supabase: search runs by meaning and by words with Russian morphology, and the results are merged. A query about Tinkoff ID finds everything on sign-in, the government services login and identification across seven years of research.
Connected with one link
A custom MCP server on Cloudflare serves the base to the agent: search, full documents, lists of files and studies. It connects with one link to Claude, Cursor, Codex and the agent in Figma. New content gets indexed with a plain git push: a GitHub Action reindexes the changes in a couple of minutes.
Rolling it out to the team
I ran a workshop for product designers: each built a prototype and added it to the team’s shared catalogue. Adding goes through a PR: the agent opens it, only a human approves. To get people started I recorded five video guides, from the first prototype to deployment.
Result
The approach is built for typical tasks: putting together a form, a widget, or adding logic to a screen. Most of a product designer’s work is like that, and that’s where the routine goes. It doesn’t replace creative or visually complex work — I still do that by hand in Figma.
- I close my typical tasks with prototypes instead of Figma — about 80% faster by my estimate
- Researchers test almost every important decision on prototypes, and the business asks to see decisions as prototypes
- Design review is moving from Figma storyboards to live flows
- I no longer check tokens and spacing in review
What’s next
Figma still holds the final screens, while prototypes carry animation, interaction and flows. The next step is for engineering to take the prototype as the final reference. The main downside is a higher barrier to entry than Figma, so I teach the team on their own tasks.
Next case
Loan application
Only half of users finished the loan form. I fit it on one screen with accordions and tested it on a vibe-coded prototype.
Banki.ruMobile web5 out of 5 for ease in testing