Luis Ouriach - How are design systems changing?

2026-03-10 - 55 min - source - Read full transcript

Vid (host)Luis Ouriach

Key insights

AI code generation flipped design systems from a nice-to-have into the literal control input for engineering output, making them more important than ever, not obsolete.

Luis says the industry swung from over-indexing on token refactoring, to briefly believing design systems were dead once AI coding tools arrived, to now recognizing that the more a team relies on a tool to generate code, the more it relies on a system to tell that tool what to code. Under-investing in the system means AI just generates 'rubbish' that piles up like a landfill for someone to fix later.

design-systems-as-ai-infrastructure

Roles are deliberately 'bleeding into' each other - designers touching code, engineers touching PRDs, PMs touching prototypes - and Luis considers this healthy rather than a threat to expertise.

He frames it as teams finally collaborating as a unit instead of passing files across silos (engineer hands off a build, PM hands off a PRD). Expertise still matters and deepens, but everyone can now contribute a rough version outside their lane, with the caveat that each role still has to hit a defined baseline output.

role-blurring-and-team-structure

Design system ownership should not live in a single siloed team, because that structure lets everyone else treat quality as 'someone else's job.'

Luis calls quality 'everyone's job' and 'everyone's business,' and argues that even the label 'design system' can be counterproductive because it invites budget and headcount conversations instead of getting treated as a thread running through everything a team builds by default.

role-blurring-and-team-structure

Writing documents before committing to pixels is still, in Luis's view, the best way to secure buy-in, even in a world where prototypes are nearly free to generate.

He pushes back on the 'prototype is the new PRD' framing: as soon as a pixel exists, people react to visual design rather than the underlying idea, so a written plan reviewed by a cross-functional team before high-fidelity work starts is still necessary to make sure the right thing gets built, even though his own instinct is to jump straight to pixels.

prototyping-and-buy-in-workflow

The access gap between enterprise and startup designers is widening because of company security policies gating AI tool access, not skill differences.

Startups get new AI tools 'out of the box' via a signed contract, while legacy enterprises may take much longer to get a tool approved. Luis calls this a 'digital divide' concern because people could get stuck in orgs that don't give them access, unable to move to more progressive ones.

enterprise-startup-tooling-gap

Once tokenized systems and AI-generated components make baseline consistency free, the strategic differentiator shifts to deliberately deviating from the system in brand-defining ways.

Luis argues an early-stage company competing on how it makes users feel may benefit from raw values and experimentation rather than premature tokenization, while a mature product scaling toward acquisitions needs the fully tokenized version; every product has to figure out where it sits on that line, which he says holds steady even as generation speed keeps increasing.

design-systems-as-ai-infrastructure

Tools like Shadcn, Tailwind, and Vercel create a network effect that is raising the baseline for what counts as clean, consistent design, making it pointless to hand-build commodity components like popovers.

Luis says accessibility and consistency that used to require dedicated advocacy inside a company can now be inherited for free from these ecosystems, so designers should stop reinventing solved problems and instead spend their effort on brand-specific differentiation that helps sell the product.

design-systems-as-ai-infrastructure

Luis's personal AI adoption moved from copying ChatGPT output into an IDE 'like a caveman' to shipping five or six Figma plugins with real users and a GitHub star, cutting build time from weeks to hours.

He describes starting in Q2 2025 using an enterprise ChatGPT license to build a plugin as an easier entry point than the React courses he had bought and never finished; iteration time compressed from about a week or two down to a couple of hours as his workflow moved into an IDE-integrated LLM tool with Git branches.

personal-ai-adoption-journey

A Claude MCP connecting Figma's canvas to a live browser lets non-technical designers test real interactions and timing, then loop back into the canvas, which Luis frames as 'not faster, but further.'

He describes pushing a design from canvas to browser, inspecting real spacing and interaction values, then returning to Figma to spin up variants without needing an engineer in the loop for days or weeks, calling the system-connected version of this loop 'the gold mine of what we've been trying to do.'

prototyping-and-buy-in-workflow

Luis is comfortable outsourcing execution to AI but flags outsourcing judgment and ideas entirely as the real risk to a designer's value.

He distinguishes between using a tool to iterate and get more sure of an idea before sharing it (which he endorses) versus 'pressing go go go all day' without engaging - the latter would leave him feeling deflated as a maker, since designers are still meant to be the experts bringing ideas to market.

role-blurring-and-team-structure

Because software is now cheap to build, success no longer requires mass-market scale - a project can be a legitimate win if it solves a real problem for a small, specific group, which Luis calls the 'niche economy.'

He cites building a Super Bowl app just for his family in about two hours, used successfully by 15 people, as proof that the old requirement (needing scale to justify the cost of building software) no longer holds; the host names this recurring idea the 'niche economy' and predicts more designers will ship many small, targeted tools rather than one big product.

personal-ai-adoption-journey

Luis rejects the narrative that AI-driven role-blurring is purely a cost-cutting tactic, arguing instead that front-end work was artificially separated from design due to past technical limits and now correctly falls back under UX.

Pushing back on a skeptical framing that businesses are just using AI to demand more for less pay, Luis argues the industry drew an artificial line between 'a picture of the front end' and the actual front end based on the technical tools available at the time, and that as those constraints dissolve, front-end implementation will be recognized as always having been part of UX design.

role-blurring-and-team-structure

Companies

Figma - Luis's employer, where he works as a design advocate; his variables launch three years earlier is used as the episode's starting reference point for how far design systems have moved.
Shadcn - Cited as a major force raising the baseline for what a clean, consistent design looks like out of the box, part of a network effect with Tailwind and Vercel that makes reinventing basic components pointless.
Lovable - Mid-episode sponsor read for an AI app builder; the newest version is pitched on autonomous planning, prompt queuing, and automated browser-based testing that catches and fixes bugs.
Granola - Sponsor segment for a meeting-notes tool; its new MCP lets the host connect AI tools directly to his meeting notes to pull topics and action items.
Inflight - Sponsor segment for the host's own feedback tool, built for sharing prototypes with video walkthroughs and collecting structured feedback.
Anthropic - Claude's MCP release for connecting design tools to code is discussed as unlocking a canvas-to-browser feedback loop for non-technical designers.

Techniques and frameworks

System as centerpiece, not a siloed team - Luis argues design systems have moved from a side function that raises quality bar to the literal input that tells AI code-generation tools what to produce, so under-investing in the system means AI generates unusable output that piles up like a landfill.
Starting-from-anywhere workflow - Luis's framing that a project can now begin in a document, a canvas, or the front end code itself, bouncing between mediums rather than following a fixed linear design process.
Document-first buy-in before high-fidelity work - Luis's personal practice of writing a document laying out findings and a plan before committing pixels, arguing that written buy-in from a design or product leader is still required before building a high-fidelity prototype, even though AI makes prototypes cheap.
Guardrails encoded as automated checks (e.g. a claude.md file) - The host describes putting rules like 'don't use raw values, use these components' into a project instructions file so an LLM enforces the design system automatically on every pull request via code review.
Deviation-from-system as the brand differentiator - Luis's argument that once tokenized systems and AI-generated baseline quality make consistency free, the strategic lever shifts to deliberately breaking from the system in specific, brand-defining ways rather than defaulting to raw values everywhere.

Summary

Luis Ouriach, a design advocate at Figma, returns to Dive Club roughly two and a half years after his last appearance to map out how design systems have been reshaped by AI code generation. He traces a pendulum swing: heavy investment in token refactoring, then a period around January 2025 when he and the host both felt design systems were becoming irrelevant as "speed" took over, followed by a correction once teams realized AI-generated code is only as consistent as the system and documentation feeding it. His central claim is that systems have moved from a siloed, quality-raising side function to the literal centerpiece of engineering work, because the more a team relies on a tool to generate code, the more it relies on a system to tell that tool what to build.

Much of the conversation explores how team structure and roles are shifting as a result. Luis describes designers touching code, engineers contributing to PRDs, and product managers reaching toward higher-fidelity prototypes, framing this "bleeding into" other roles as a positive return to teams collaborating as a unit rather than passing work through silos. He is careful to note this doesn't erase expertise - each role still has a baseline expected output - but argues organizations should stop treating the design system as a corner team's job and instead treat quality as everyone's responsibility, even avoiding the "design system" label at times because it invites unwanted budget and hiring conversations.

On workflow, Luis pushes back gently on the idea that prototypes have fully replaced PRDs. His own practice is still to write a document laying out findings and a plan before committing to pixels, because once a visual exists people react to the visual rather than the underlying idea, and getting sign-off from a design or product leader beforehand remains valuable even as prototyping gets cheaper. He and the host also discuss a hands-on example of guardrails: encoding rules like "don't use raw values" into a project instructions file so an LLM enforces the system automatically during code review, and the tension between over-tokenizing (robust but inflexible) versus intentionally allowing raw-value deviation as a brand differentiator once AI makes baseline consistency nearly free (with Shadcn, Tailwind, and Vercel cited as forces raising that baseline for everyone).

The episode also covers the widening gap between enterprise and startup teams, driven less by skill than by which companies' security policies let designers actually get AI tools approved, and Luis's own adoption journey - from copying ChatGPT output into an IDE "like a caveman" to shipping five or six Figma plugins with real users and GitHub stars over the past year. He describes a Claude MCP release that connects Figma's canvas to a live browser as unlocking a much faster idea-to-feedback loop, especially for non-technical designers, and reflects that outsourcing craft entirely to a tool - rather than using it to get more sure of an idea before sharing it - is the outcome he actually worries about. The conversation closes on career reflections: Luis says he no longer tries to plan his career or keep up with every new tool, focusing instead on writing, visual design quality, and community, and pushes back on the "AI is being used to justify smaller teams and lower pay" narrative, arguing instead that the true front-end/UX line has just been artificially split by tooling limits that AI is now dissolving.

Notable Quotes

"The more we rely on a tool to create our code, the more we rely on a system to tell that tool what to code." - Luis Ouriach

"I thought we are all toasted. We are in trouble here because our jobs are no longer relevant. And that was just a thankfully temporary feeling." - Luis Ouriach

"A design system is a bar of quality in an organization. And that's everyone's job." - Luis Ouriach

"Writing down before you commit to a pixel is the best way to get buy-in for an idea." - Luis Ouriach

"Outsourcing everything you do to something is where I get a little bit more concerned... we are the experts in seat to bring something to market." - Luis Ouriach