I have a thesis that I am not fully comfortable with yet:
Design is the intermediate representation of software. As AI learns to generate software directly, the value of design artifacts gets thinner.
That statement sounds like a dismissal of design. It is not meant to be. It is a claim about where value sits in the chain, and how it shifts when generation becomes cheap.
I might be wrong. But the recent wave of AI tooling makes the pattern feel real enough to write down.
What I mean by "design"
When people say "design," they often mean different things. I think of it as two layers:
- Design as artifact: screens, flows, prototypes, component specs, UI systems.
- Design as judgment: the ability to define the problem, set constraints, and decide what "good" feels like.
The first layer is tangible. It becomes files and links and handoffs. The second layer is more invisible. It lives in the decisions you ship and the standards you keep.
If my thesis is true, it only applies to the first layer. It is not saying judgment is less valuable. It is saying the artifact is no longer the main bottleneck.
Why AI makes the intermediate layer thinner
Software creation used to be a staircase:
- idea
- design
- implementation
- iteration
Design was a crucial intermediate representation. It reduced ambiguity, aligned teams, and translated intent into something buildable. It was the safe place to explore before you paid the cost of engineering.
AI collapses the staircase. It turns the pipeline into a loop:
- idea
- generation
- evaluation
- correction
If you can go directly from intent to working code, the need for a heavy intermediate layer declines. You still need clarity and constraints, but you do not need as many pixels to get there.
So the value shifts from "make the thing" to "define the thing." That is the crux.
Recent practice: AI operating the design stack
This is no longer theoretical. We are already watching AI operate the design stack directly.
Two examples that matter:
- Claude Code operating Figma: people can now instruct an AI agent to open Figma, manipulate frames, adjust components, and update layouts. The UI becomes another API surface.
- pencil.dev with MCP: tool providers expose design operations through MCP, so an AI can create screens, restructure layouts, and apply design system components with structured commands.
What is new here is not "AI can make pretty pictures." It is that AI can act inside the actual production tools.
That changes three things:
- Capability (what AI can do inside the tool)
- assemble UI from design system components
- align spacing, typography, and color tokens
- refactor layouts across multiple screens
- Workflow (how teams move work forward)
- shorter handoffs between design and engineering
- faster iteration loops with fewer blocking artifacts
- more "live" design changes late in the process
- Evaluation (what becomes the bottleneck)
- quality is limited by the clarity of intent
- standards and constraints define the output quality
- teams need better review systems, not just better mocks
This is why I think the artifact layer is thinning. Not because design is unimportant, but because AI is now doing the production work inside the same tools designers use.
What actually gets thinner
If the pipeline collapses, the things that get thinner are mostly delivery artifacts:
- pixel-perfect mocks for every edge case
- long UI specs that can be inferred from tokens
- static flows that are already outdated by the time they are reviewed
- handoff documents that exist only because the next step cannot move without them
Those artifacts were necessary when the cost of implementation was high and slow. They are less necessary when an AI can generate a working prototype in minutes and keep it aligned with the design system.
This is the part of design that looks like a middle layer. It is the "map" you needed to hand to engineers so they could build the territory.
If the territory can be generated directly, the map gets thinner.
What gets thicker instead
The thinning of artifacts makes other forms of design more valuable, not less.
What gets thicker:
- Problem definition: what is the user actually trying to do, and why does it matter now?
- Constraint setting: what must be true for this product to be safe, coherent, and trustworthy?
- Quality judgment: what does "good" feel like, and what will we refuse to ship?
- Narrative framing: what is the story the product tells, and how does it earn trust over time?
These are not easy to automate because they require context, stakes, and taste. AI can propose options, but it cannot own the consequences.
So the center of gravity moves from outputs to decisions.
Where the thesis breaks
There are places where design artifacts remain heavy and will stay that way.
- High-risk domains: healthcare, finance, civic systems. The cost of mistakes is too high, so the intermediate layer must be explicit and auditable.
- Complex stakeholder environments: when many parties need alignment, design artifacts become the contract.
- Brand and differentiation: when the product must feel distinct, the artifact itself carries the identity.
These are the places where design artifacts are not just a proxy for intent; they are the final product.
So the thesis is not universal. It is conditional: the lower the risk and the more commoditized the interface, the thinner the artifacts become.
The new design role: intent architecture
If the artifact layer thins, designers do not disappear. They change shape.
I think the most valuable design work moves into three roles:
- Intent architect: define what the system should do, not just how it looks.
- Evaluator: build the standards and tests that keep AI output aligned with user needs.
- System gardener: maintain the design system and its values as the product evolves.
In other words, designers become the people who make the AI productive, safe, and coherent. They are less the authors of static screens and more the stewards of product direction.
This is a different kind of craft. It is closer to product management and closer to engineering. It is also closer to judgment.
A better version of the claim
So is the original claim reasonable? I think it is, but it needs a cleaner phrasing.
A sharper version might be:
Design artifacts are getting thinner as AI learns to generate software directly, while design judgment becomes more valuable because it defines what AI should build and how we decide if it is any good.
That is less provocative, but more accurate. It does not diminish design. It explains a shift in where the leverage lives.
Closing thought
When generation becomes cheap, the artifact is no longer the moat. The moat is the standards you set and the taste you enforce.