AINEWS
One-shot AI design tools are solving the wrong problem
The emerging 'loopmaxxing' vocabulary in AI design tools argues that iterative human steering — not one shot generation — is the right frame for evaluating design agents.
The emerging 'loopmaxxing' vocabulary in AI design tools argues that iterative human steering — not one shot generation — is the right frame for evaluating design agents.
The dominant pitch for AI design tools is unambiguous: replace the human, get the finished interface in one shot. An open-source practitioner has been making a quieter counter-case, arguing that the goal itself is wrong, and that the tools that matter will keep the human steering.
Paul Bakaus, a designer and developer, built Impeccable, an open-source collection of design "skills" for AI coding agents such as Cursor. He presented the framing at the AI Engineer World's Fair, a conference for engineers building AI agent systems, and detailed it on the AI engineering podcast Latent Space. His argument is not that AI is bad at design. It is that one-shot generation is the wrong target, and design tools deserve to be evaluated by how well they help a human direct the work.
Bakaus calls the discipline around this kind of work "skill engineering." The vocabulary is deliberately small. The Impeccable skills name specific design moves a reviewer might call out in a critique: "bolder," "quieter," "denser," "more polished." Each skill is a unit of context an agent can apply to the work in front of it, scoped tightly enough that a designer can reason about what changed and steer the next pass.
The intended user experience is iterative. "The point is to give you a way to steer what you want to end up with," Bakaus said in the Latent Space interview. "It's never going to be a tool for one-shot design. That's not the intent."
Headline demos in this category have tended to be one-shot: paste a wireframe or screenshot, get a redesigned screen. The implied future is one in which the autonomous model replaces the senior designer. Bakaus's position is that the senior designer is the point. Useful agent output, as he frames it, requires domain knowledge, contextual cues, and an explicit human steering handle. The agent carries the first two. The human keeps the third.
He labels this pattern "loopmaxxing": humans remaining inside iterative steering loops with the agent, rather than handing the loop to the agent and walking away. The label is Bakaus's. What it names is an older practice in other creative fields, where a practitioner marks up drafts and asks for another pass.
There are reasons to be cautious about how universal this framing is. Bakaus is the creator of the project whose design philosophy he is articulating, and the Latent Space conversation and the AI Engineer World's Fair talk are built around what Impeccable is meant to be. The dispute he sketches, between tools built to replace designers and tools built to be steered by them, is plausible, but the public reporting on hand here is one creator's case. Whether major AI design platforms end up treating skills like Impeccable as first-class input, or absorb the equivalent into their own systems, is a question that needs independent practitioner reporting to settle.
What is worth tracking next is whether coding-agent platforms, including Cursor and broader agent runtimes, expose explicit hooks for skills and spec layers as separate artifacts. Bakaus's argument predicts they will. The IDE and agent product roadmaps will show it either way.