I went into a recent project excited. The newer LLM tools had crossed a threshold for me. They felt like something I could finally lean on hard for a tricky visual-design problem: cross-domain research, concept generation, and prototyping all running in one loop. I was genuinely confident going in.
And it delivered. The LLM helped immensely. But it still wasn’t an ask-and-be-done problem.
I had spent two weeks generating comparison concepts with the model — gradient overlays, ISO lines, diff heatmaps, ghosting animations, side-by-side stacks — when a teammate said one sentence in a meeting that collapsed the whole problem into something that unlocked our final approach. The AI hadn’t been wrong. It just couldn’t have asked the question that mattered.
That gap between the work the LLM did brilliantly and the work it could not do is what I think is important.
The Problem, in the Abstract
The design problem was a side-by-side comparison of two complex spatial signatures from the same person doing the same activity. Dense data living on a 3D surface, changing over time. The user already has a reference in their head — “this is what good looks like for me” — and they want to know how their latest attempt stacks up against it.
“Just put them next to each other” sounds easy. It isn’t. Occlusion, color perception, no shared axes, no obvious hierarchy, no clear direction of difference, single-frame views that hide everything happening over time. The version we had partially worked, but it was actively misleading on at least three axes.
This is the kind of problem I used to spend a long time alone with a sketchpad on. This time I had an LLM in the loop from day one. Here’s what changed, and what didn’t.
Where the LLM Was a Force Multiplier
Cross-industry research at a depth I would never have done alone. I asked the model to pull every documented approach to “comparing two spatial signatures” across radiology, GIS, mechanical tolerancing, sports broadcast, weather, and A/B testing dashboards. Each industry handed me a vocabulary — anti-patterns to avoid, studies on color perception, known failure modes. By the end of the second day I had about fifteen viable directions and a list of dead ends, before I’d drawn a thing. That kind of breadth used to take me weeks of skim-reading and feel half-finished. The LLM finished it in an afternoon and cited as it went.
Framing the failure modes of what we had. I pointed the model at our existing baseline and asked it to enumerate what was wrong. It came back with a numbered list of seven specific failures: occlusion, lack of direction, time-collapse, colorblind failure modes, and a few others. That list became the rubric every new concept had to clear. I’m slow at structured critique on my own. The LLM is fast.
Generating and prototyping concepts at a speed I’ve never had. We produced about eight distinct concept directions, each with a working prototype, in roughly three days. Storybook scaffolding meant we could pick from variants in a working meeting instead of debating them in the abstract. This turned out to matter more than any single concept. The ability to diverge into any number of concepts and bring back the best ideas of each meant we could chip away at the solution much faster than before.
Where the LLM Fell Short
I want to be honest about this part, because the breathless takes are everywhere.
It Couldn’t Deliver the Final Version. We Still Had to Iterate, a Lot.
I can count the visible iterations in our sandbox. Seven superseded baseline builds. Eight idiom explorations — “what if it looked like topography? what if it looked like a radar sweep? what if we used hard borders instead of gradients?” Then a fresh research and adversarial-verification pass produced four more candidate specs. Every iteration solved one thing and broke another. The LLM was great at “give me ten more variations.” Weak at “now make the right one.”
The Breakthrough Was a Human Reframe.
What unlocked the design was a single sentence from a teammate who actually does the activity we were designing for: “I just want my reference on the left and the one I just did on the right.” We got to the why behind it.
That sentence dissolved most of what we had been chasing as the centerpiece of the visualization. Two heatmaps, layered or paired, asymmetric color so the foreground reads against a muted base. The diff calculation is still there — it surfaces as a similarity score, a single number that flags whether the visual is worth digging into. But the visual itself doesn’t have to carry the math. The user’s eye does most of the comparison work, and the score handles the rest.
The LLM could not have proposed this on its own — not because the model isn’t capable of it, but because it was answering the variants of the question I had asked: what’s a good way to compare two signatures? The right question was a different one entirely: what does the practitioner ask themselves when they look at this? Its unlikely just prompt engineering alone would have gotten us a better result than having the person who does the activity in the room.
Pure Prompting Was Slow. Diverging and Converging With the Team Is What Got Us There.
I want to be clear-eyed: I believe the model could have arrived at the simpler frame. I believe prompting alone was the slow path to it. I was asking the wrong question of the model the same way I was asking the wrong question of the design — and no number of follow-up prompts was going to surface that on its own. I had to go through the learning journey to be able to understand the problem to ask the right question. What did surface it was opening the loop: diverging with the LLM on research and concept generation, then converging with the team and the practitioner-experts in the room. That alternation — wide with the model, sharp with the humans — is what produced a solid solution. The prompts were a multiplier on the team process, not a substitute for it.
The Recipe I’ll Run Next Time
When I hit a problem like this again, here is how I will sequence it.
- Diverge wider on research, first. Use the LLM aggressively for cross-domain patterns before sketching anything. Force yourself to leave the domain.
- Move to visuals early. I’m a visual learner. Concepts only register for me once I can see them. The visuals also help me find the right problem statement — they shouldn’t wait for it.
- Generate three to six concepts. Not twenty. Not one. Six is enough to surface the strong moves and the obvious dead ends. More than that and you’re researching, not designing.
- Now frame the actual problem. With visuals on the table, sit with the practitioner and ask: what question are you trying to answer when you look at this? That answer reframes the problem, every time.
- Converge by mixing. Keep the moves that work, discard the rest, custom-build from the building blocks. The final design is rarely one of the original six — it’s a remix.
- Validate. Put it in front of users. The LLM cannot tell you whether they’ll get it.
The Takeaway
The LLM is the best research partner I have ever worked with. It is also the worst final-design collaborator I have ever worked with. Knowing where the line sits between those two roles is the actual skill. If you’ve been holding off on using AI for design because you’ve heard “it can’t really design” — you’re right, and you’re missing the point. Use it for divergence, research, and iteration speed. Protect the reframe, the synthesis, and the human in the room. You’ll move a lot faster on the work that actually matters.