AI prototyping kit

ClaudeFigma
2026

Banki.ruDesign system for vibe coding

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.

The canvas: live prototype screens in phone and desktop frames

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.

Diagram: the kit, canvas and knowledge base around the agent

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.

The agent spins up a prototype: name question, install, link

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.

Illustrations page in the kit: each one says when to use it

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.

State configurator of an organism: signing for a microloan, loan and deposit

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
A comment on a button on the canvas and what the agent gets from it

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.

A task card in the knowledge base: current decision, open questions, meeting history

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
A research slide and the same slide in markdown

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.

Diagram: Obsidian → GitHub → Supabase → MCP → Claude, Cursor, Figma

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.

The Figma agent finds research for a mock-up and cites the source

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.

The team’s shared prototype catalogue

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

ClaudeFigma
2025

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