(Re) Designing A Design System for LLMs
Prism is Kinective's design system. The first version was a Figma library built by hand. It had 815 variables, dark and high-contrast themes and kiosk typography, and using it correctly meant knowing how it had been built. Designers outside the core team drifted from it, and when the company started prototyping with AI, there was nothing an AI tool could read.
Prism 2.0 is the rebuild. Figma is the source for how things look, the code repository is the source for how things behave and what accessibility requires, and a build turns both into the packages and documents that designers, engineers and AI coding agents all use. Automated checks flag any mismatch between the two.
- Role
- Head of UX, founder and owner
- Team
- My 3 designers, 2 FE devs
- Stakeholders
- VPs of Product, Engineering, Marketing
- Scope
- 5 product lines, 10 products
- Frameworks
- React and Vue
- Timeline
- 6 weeks, Aug 19 to Sep 28, 2026


Key figures
100
Components with a machine-readable accessibility spec
Agent
Ready9
Drift checks between Figma and code, run on every build
Overview
Project overview
Prism is the design system under Kinective's 10 banking products. I founded it and I own it. Prism 1.0 did its job for the people who built it, and by mid-2026 two things made that untenable: product files were drifting from the library, and AI tools could not read it at all.
TL;DR
- Prism 1.0 was hard to adopt, and AI tools couldn't read it. AI-built prototypes made up their own UI, and designers had to rebuild them.
- We rebuilt Prism in six weeks. Figma defines how things look, the code defines behavior and accessibility, a build generates the packages and docs from both, and automated checks flag any mismatch.
- AI agents read one file to learn every Prism name, look up details through an MCP server, and have their output checked automatically.
- Accessibility requirements are written for all 100 components, so they're built in from the start.
- Status: the foundation is done and components are still being built. Two automated checks and one contrast issue are open.
Measured in the files
1
File that teaches an agent all of Prism, in about 8,000 tokens
0
Semantic tokens holding a raw value, down from 106
100%
Components with a machine-readable accessibility spec, up from 0
4
Devices from one component set, on one attribute
On a test file of known mistakes, the output checker now finds 22 problems in AI output where it used to find 4.
Impact
Some gains are measured in the files today. The rest come from the pilot. The outside studies we planned against are at the end, labelled as theirs.
Prism 1.0 and 2.0 compared
100%
Semantic tokens that alias a primitive, up from 77%
95%
Variables scoped to specific properties, up from 28%
100%
Components with a machine-readable accessibility spec, up from 0
Measured
- What an agent loads to know Prism: Nothing reliable, explained again in each prompt → one file, about 8,000 tokens, the full vocabulary
- Problems caught in AI output: 4 on a violation probe → 22. On a sample Prism prototype, 28 → 71
- Component sets needed per device: Per-theme and per-size copies on the page → one set, four modes by attribute
- Re-syncing Field after an Input change: Hand work, variant by variant → one scripted mirror pass, verified 20 of 20
- First fully AI-built, Code Connected component: Avatar for Vue, built by the pilot team on September 25
The prototype workflow
A stakeholder AI prototype, before
- A leader prompts a working UI in an afternoon.
- It arrives looking finished; people have seen it work.
- Design works out what the invented UI was meant to do.
- Design rebuilds it on brand, accessible, and buildable in React and Vue.
- Engineering rebuilds it again from scratch.
The same prototype on Prism 2.0
- The agent loads
design.mdand the Prism MCP. - It builds from real components and role tokens.
- Automated checks catch problems before anyone reviews it.
- Design reviews the experience and the flow.
- Engineering builds on the same code.
Why this works
Engineering
47%
faster to build a form with a mature design system than from scratch, including the time to learn it.
Source: Sparkbox, "The Value of Design Systems Study"
Design
34%
faster task completion for designers working with a relevant design system.
Source: Figma, "Measuring the value of design systems"
That's the short version.
The full study goes through the problem in detail, the six decisions and what each one cost, how the build ran, what went wrong, and what I'd do differently.
Hey there,Thanks so much for checking out my portfolio! I really appreciate you taking the time to look through my work.Best,




Start the conversation
Whatever you're working on, I'd like to hear about it.













