Ask an AI to design something. What do you get?
Usually: technically correct but aesthetically generic. The spacing is fine. The colors are acceptable. The typography doesn't embarrass anyone. But it looks like it was designed by someone who studied design from textbooks and never actually had taste.
Impeccable is trying to fix that.
Why AI Design Has Been Bad
The core problem isn't capability — it's training signal.
Language models are trained to predict text. Visual quality is hard to capture in text. A model can learn that "good design has sufficient whitespace" as a linguistic pattern without understanding what sufficient whitespace actually looks like and why it matters.
Design also requires a kind of visual memory — knowing that this shade of blue feels cold and corporate, or that this font combination has been overused to the point of becoming a signal for lazy work. That knowledge is embodied in examples and context, not in design principles written as text.
What Impeccable Does Differently
Impeccable approaches this as a tooling problem, not a model problem. Rather than trying to train a better design model, it gives existing AI models better design infrastructure:
Design system access: Impeccable connects to established design systems (Tailwind, Material, your custom system) and constrains AI suggestions to tokens and components that already work together.
Visual feedback loops: Rather than generating a design and calling it done, Impeccable renders the output, takes a screenshot, and passes it back to the model for iteration. The model can see what it made.
Constraint-based generation: Instead of open-ended "design this," Impeccable prompts the model with specific constraints — spacing scale, type scale, color palette — that guide output toward coherent visual systems.
impeccable generate \
--component "pricing card" \
--system tailwind \
--brand-colors "#2563eb,#1e40af" \
--iterate 3
The --iterate 3 flag means it renders, screenshots, evaluates, and regenerates three times before returning output.
The Visual Feedback Loop
The screenshot-and-iterate approach is the genuinely clever part. It's a simple idea that has a large impact.
Without it: the model generates HTML/CSS, you see the output, you prompt again. The model can't see what you see.
With it: the model generates HTML/CSS, the tool renders it, takes a screenshot, sends the screenshot back to a vision-capable model, asks "what's wrong with this design?", gets back specific critiques ("the text contrast is too low," "the button is too small relative to the card"), and incorporates them in the next generation.
It's implementing a design review loop automatically.
What It Actually Produces
Honest assessment: Impeccable produces better AI-generated design than naive prompting, but it's not a replacement for a designer.
Where it's clearly better:
- Component-level UI (cards, buttons, forms, modals)
- Layouts that follow established patterns (landing pages, dashboards)
- Matching an existing design system
Where it still struggles:
- Brand identity and personality — things that feel distinctive rather than competent
- Motion and interaction design
- Novel layouts that break from convention intentionally
The Practical Takeaway
For developer-built tools, side projects, and internal tooling — where "competent and consistent" is the real bar — Impeccable is useful. You'll get results significantly better than "AI generated this" without needing to hire a designer.
For consumer products where visual differentiation matters — you still want a real designer. Use Impeccable for the scaffolding, not the final artifact.
The right mental model: Impeccable is to design what linters are to code. It catches the obvious problems and enforces consistency. It doesn't replace the engineer who makes the architecture decisions.