Why AI Can’t One-Shot Design and What That Means for Your Team

Paul Bakaus argues that AI cannot one-shot good design. Direct manipulation and fully agentic approaches both fail—the former is too slow, the latter produces generic slop. The middle ground is using adjectives and verbs as a shared language to steer AI output iteratively.

His tool, Impeccable, encodes design commands like "bold" or "quiet" with project-specific meaning, giving engineers and designers a way to guide AI without micromanaging pixels. This approach rejects the popular demand for auto-mode, insisting that taste and context require human judgment.

For teams, this means rethinking workflow: design is not linear. Injecting human feedback at the right altitude—through vocabulary, not prompts—can improve output quality. The tradeoff is that teams must invest in building shared design language and accept that the last 20% of polish still needs human hands.

Hype? The claim that adjectives alone solve the control problem is unproven at scale. Evidence is anecdotal, based on Bakaus’s own builds and community feedback. Engineering leaders should watch for whether this pattern generalizes beyond early adopters.