Here's the uncomfortable truth about AI-generated design work: the robot is not the problem. You are. Or more precisely, the junk drawer you handed it is.
Bad naming conventions. Spacing values pulled from thin air. Figma components that got detached six months ago and nobody noticed. Decisions that lived in someone's head and never made it into a single document. Then everyone acts shocked when the AI spits out oatmeal.
You gave it vibes, a screenshot, and a prayer. What did you expect?
The Quiet Part the Industry Just Noticed
A sharp piece by Hardik Pandya of Atlassian — surfaced by Smashing Magazine — makes the case for building "AI-ready" design systems. The core idea: AI-generated prototypes fail not because AI is broken, but because the design systems feeding them are riddled with tiny inconsistencies. Hard-coded values. Decisions made but never documented. Components that drifted from their definitions.
The fix isn't a better AI. It's better human groundwork.
I agree with all of it. But I want to say the part that gets glossed over in the polite industry write-up: designers have been ignoring design-system hygiene for years because humans could fudge the gaps. A skilled developer could look at a messy Figma file and intuit what you meant. A junior designer could ask. A client would never know.
AI cannot fudge institutional mess gracefully. It will either hallucinate around the holes or faithfully amplify the chaos. Five seconds of generation time, and suddenly every bad decision you deferred is right there on the screen in production-ready code.
So yes, now everyone is discovering that documentation matters. Welcome to 2014.
Design Decisions Are Infrastructure. Full Stop.
This is the line from the source piece that actually matters:
Every design decision — not just what something looks like, but how you prioritize and why — needs to find its way into a spec file that AI can read.
That is dead-on. AI does not need more LinkedIn motivation. It needs rules. Named tokens. Component logic. A clear "do this, not that" file written in plain Markdown that it can consume every single time it generates something.
The practical toolkit here is straightforward:
- Spec files — structured Markdown documents with spacing rules, color choices, component usage guidelines, and priority logic. Not a mood board. A rulebook.
- Token layers — a closed set of named variables (spacing, color, type scale) that AI always pulls from, so it stops inventing plausible-sounding values from scratch.
- Audit scripts — automated scans that catch hard-coded values and flag anything that's drifted from the spec. FigmaLint is a free Figma plugin that does exactly this: catches detached instances, missing interactive states, hard-coded colors, bad token bindings.
- Sync routines — when the design system updates, a process that flags which spec files are now stale. The AI should always be reading current rules, not month-old ones.
None of this is exotic. It's just discipline applied to a system that most teams let slide because the stakes felt low. The stakes are not low anymore.
What This Actually Looks Like in Practice
Let me make this concrete, because "improve your design system" is advice that helps nobody.
At JMS, I have a small but controlled production lane for content. Article draft. A header image system built around a specific set of rules: dark, cinematic backgrounds, low visual clutter, built to sit behind a defined typography system without wrecking readability. The AI doesn't get to make taste decisions. It gets constraints. Named parameters. A spec.
The result isn't "AI made this." It's "I used AI as a production engine inside a real design system." That distinction matters enormously — for quality, for consistency, and for the sanity of anyone who has to maintain it six months from now.
That's the model worth stealing. Not "AI, run my business." A contained assembly line with human taste gates at every approval point.
The Real Tradeoff Nobody Mentions
Building this properly takes time upfront. Writing good spec files, auditing your tokens, cleaning up years of accumulated Figma chaos — that's real work. It's not a weekend project.
And if your business is small and your design needs are simple, you may not need a full Atlassian-grade design system. What you do need is some version of this thinking: documented decisions, consistent naming, and at least one set of rules the AI can follow instead of guess.
The tradeoff is time now versus chaos forever. Every shortcut you skip today becomes a hallucination you explain to a client tomorrow.
The One Thing to Take Away
Clean your room before you ask the robot to build an extension.
That means: before you invest another dollar in AI tooling, audit what you're actually feeding it. Are your design decisions written down anywhere? Are your spacing values named or just "eyeballed in Figma"? Do you have a single document that explains how your brand actually works?
If the answer is no, the AI isn't your problem. The junk drawer is.
Fix the junk drawer first. The automation gets genuinely useful fast — but only once you've built a system worth automating.
—
If you're trying to figure out what a sane, documented, AI-assisted workflow actually looks like for a small business — not a 200-person product team — let's talk. That's the exact problem I spend my days on.