Context
AI systems generate slide decks and design artifacts constantly now. Getting them to generate good ones requires someone who can say, precisely and repeatably, what separates a strong result from a plausible one. That precision comes from human designers. My work at Mercor sits there. I evaluate AI-generated slide decks and design outputs, and I write the structured critiques that explain each judgment.
Problem
"Good design" is mostly tacit knowledge. A designer sees a slide and knows in seconds that the hierarchy is off. Left as a reaction, that teaches a model nothing. The work is making it explicit: which principle, where, and what the fix would be. Design quality also runs on several axes at once. The same deck can be beautiful and empty, or follow every instruction and still be ugly. So a single verdict is usually the wrong output.
What I did
I evaluated more than 500 AI-generated decks, websites, and other design outputs across six evaluation frameworks, each with its own criteria. In my own terms, they split quality into separable questions:
- Aesthetic rankingVisual quality alone: composition, hierarchy, color, typography.
- Style matchingHow faithfully an output takes on a target style without sliding into pastiche.
- Content gradingWhether the substance holds up: accuracy, structure, the right weight per slide.
- Template creationWhether a generated template is genuinely reusable. Consistent, flexible, systematized.
- Template followingWhether new content respects an existing system or quietly breaks it.
- Instruction adherenceStrict conformance to written direction. Deviation is the failure, regardless of taste.
The axes ask for different postures. Aesthetic ranking wants calibrated taste. Instruction adherence wants the opposite, which is the discipline to score against the brief when your taste objects. I switch between them many times a day, and the switching is most of the difficulty.
Each evaluation came with a structured critique: the issue, the principle behind it, the evidence in the artifact, and what a correct version would do. I have written more than a thousand of them.
Each critique names the issue, the principle behind it, the evidence in the artifact, and what a correct version would do instead.
Method · six evaluation frameworks
The six frameworks and the different judgment each one asks for. Opens full size.
What I found
Two things. Writing down why something works sharpens your eye faster than making things does. You can't hide behind intuition once every judgment has to be defended in writing. The second thing is about the outputs themselves. The most common failure turned out to be a kind of confident indifference to context. The work looks finished. It just ignores what this particular content needed. I have written that critique, in specific terms, many times.
What changed
The critiques feed into improving the quality and human-centeredness of AI design output. The habit transferred, too. I now review my own design, film, and audio work the same way, in writing, before anyone else sees it.
The evaluation work is described here in my own words, at the level of method. Details of client materials stay private.