A new foreword: the agentic AI update

When I started writing this methodology in 2020, I was expecting the concept of design systems to mature and consolidate over the course of the decade, making it easier to grasp and delineate from neighboring concepts. Six years later, only half of that has come true. Design systems have certainly matured, but instead of consolidating, the notion has kept expanding — most recently (and most forcefully) with the rise of agentic AI. Where does the system live, how is it documented, how is it governed, consumed, communicated? What are the tools we should be learning and using? And what will all of this look like five years from now? Or even in just two? I feel like I have more questions than I had six years ago.

Then there are the agents themselves: how do we teach them, how do we instruct them, and how do we follow what they are doing to our artifacts and our products? For us as designers, front-end developers, and product managers, these are scary times. But they’re also exciting and, honestly, a lot of fun, too. The road is just as winding as before, but we have been handed the keys to a much faster car. And we definitely want to avoid driving into a wall at speed.

Lean Design System is receiving a long-overdue refresh, five years after its initial publication in 2021. In this new foreword, I want to summarize what has changed and why.

Collaborating with AI

The first time I used agentic AI to scaffold a website project from scratch, I had two immediate thoughts: that this was remarkable, and that Lean Design System had just become obsolete. All the front-end knowledge collected over the years (the methodologies, the naming conventions, the design system practices) seemed to have lost its purpose in an afternoon.

However, the more I worked with tools like Claude Code and Cline, the more convinced I became of the opposite. Workflows, timelines, and tooling do need reevaluating, urgently. But the core ideas and motivations behind the methodology still hold. The native material of digital products is code, and codified design allows us to automate, scale, and share ownership. But even more importantly, codified design is the foundational shared language between human designers, developers, and our new agentic coworkers. Codified design maps the codebase to our design decisions, and it allows us to communicate our goals and ideas precisely. Now that generating new components is nearly free, design debt can accumulate much more quickly. But agentic AI also enables us to audit, clean up, and prune the design system efficiently, so that it stays lean and focused on what’s important — which in turn is essential for it to serve its purpose and make our lives easier, not harder.

Here is a recent example from my daily work: when creating new components or widgets, I will always define a BEM (Block-Element-Modifier) structure first. Naming every piece of a component forces me to make my mental model explicit, and that process often reveals flaws and discrepancies. Sometimes Claude flags a weakness in my proposed anatomy, and we end up with a better solution. Sometimes I realize it misunderstood the bigger picture, and we can backtrack to where the misunderstanding started. The methodology becomes a shared language between me and the AI. Compared to simply delegating, a methodology-driven approach feels more like working in tandem with a very fast colleague. Structured approaches like this allow us to build products with AI that are maintainable, reliable, and visually consistent.

Kent Beck, the creator of Extreme Programming, wrote that AI had dropped the value of 90 percent of his skills to zero, while the leverage of the remaining 10 percent had gone up a thousandfold.¹ I’m not sure about the exact ratio, but I agree with the general idea: understanding best practices and proven methodologies is not only still relevant but, I think, a requirement for collaborating productively with agentic AI. These frameworks were designed to help humans collaborate, but they can do the same for humans and AI, in the sense that they provide common ground where both sides understand the rules.

(There is a certain symmetry in Beck being the one to say this. Extreme Programming’s structure of individually usable practices is the template the 12 tactics were built on.² And it was Beck’s openness to vibe coding that inspired me to experiment with agentic AI in the first place.)

You have been promoted

Modern LLMs — and agentic AI in particular — fundamentally change how we work on and with design systems. We can now explore, experiment, build, audit, and deploy so much faster than before that we are forced to push our ambitions further. Things that seemed impossible just a while ago suddenly deserve a second look.

A handful of examples off the top of my head: if you are designing for a fast-moving fashion brand, you can now take bigger swings. Why not have a seasonal design system, based solely on the core design principles of your brand — automatically rebuilt each cycle instead of being maintained manually? If you work in SaaS or B2B, your design system could plug directly into an AI layer. You control all the design atoms, and your users just prompt any dashboard they need — generated straight from the building blocks of your library. Are you dealing with a messy and outdated codebase, or a stale UI kit that has been drifting from your production builds for months? Redeveloping your design system from the ground up is now a matter of days (or even just hours) instead of weeks or months.

Whatever may have seemed too risky, too expensive, or too daring two years ago has now become feasible, maybe trivial. And whatever your role was before, you are now a team or product lead. Congratulations! This promotion comes with all the benefits of delegating work, but also with new responsibilities: to make the best use of your new team, you have to gain a high-level understanding of new concepts, dependencies, and best practices. At least today, your agents still require solid and coherent plans and clear instructions — someone who understands the big picture and can help get a botched branch back on track.

The success of a design system isn’t defined by its comprehensiveness or depth of implementation, but solely by the extent to which it helps us solve real problems and generate compelling products. This core tenet of the methodology was an important corrective back when components were expensive, but now that producing them costs almost nothing, it has more or less become the whole game.

Changelog: what’s in the update

With this update, all tactics and diagrams will be overhauled. Changes include the correction of outdated references, examples, and best practices, and also general improvements with the goal of making the methodology much more accessible and highlighting the opportunities that agentic AI opens up for us.

Agentic AI features: llms.txt, AGENTS.md, skills

A newly added llms.txt and an AGENTS.md file make the methodology accessible to AI crawlers and agents. Some tactics will also provide skills that can be installed in harnesses like Claude Code or Cline. The skills will be explained in the corresponding articles. Once these skills reach a certain level of maturity, I will consider packaging and providing them as installable plugins.

The infrastructure layer will be renamed to groundwork

I was never quite happy with the naming of the infrastructure layer. It always felt too clunky and even potentially misleading, as infrastructure is already a very loaded term in software development. For these reasons, the infrastructure layer will be renamed to groundwork, which captures the idea behind the layer much better. Tactics in the groundwork layer describe what helps you build a solid foundation for your design system work.

The survey section was removed

So much has changed since 2021 that I decided to remove the survey section entirely. It just didn’t seem relevant any longer. Maybe I’ll set up and conduct more surveys in the future, but for now, surveys are gone.

Introduction and link to the 2021 thesis paper

Lean Design System was originally based on a master’s thesis of the same name. I wrote it at Hochschule für Gestaltung (HfG) Schwäbisch Gmünd in 2021. Many thoughts and conclusions described in it are outdated today, and it’s only available in German — but it’s still the origin of all the core motivations and ideas of the methodology. So I will add a link to it, mostly for completeness sake.

Kent Beck, 90% of My Skills Are Now Worth $0, 2023. ↩
Kent Beck, Extreme Programming Explained, 2nd ed., 2004. ↩