I’ve watched a lot of AI content over the past year.
Most of it is about going faster. Automating more. Getting the machine to do the thing so you don’t have to.
This video from Dan Hollick is different.
Dan is an illustrator working on a technical book. AI is typically terrible at technical illustrations that need to be accurate. So instead of asking AI to make his illustrations, he asks AI to help him make tools that help him make illustrations.
The distinction matters more than it sounds.
Every tool has a resolution
Dan opens with a mental model I hadn’t heard articulated this clearly before.
Every tool imposes a resolution on the work. A generic tool like Figma has low resolution. You get there in fewer steps, but you can’t get exactly where you need to go. A specialized tool has high resolution: more steps, but you can get much closer to the final result.
Here’s how he put it:
“When you use a generic tool, the resolution is quite low, which is cool because it doesn’t take that many steps to get where you need to go, but you can’t get exactly where you need to go. And then when you use a very high resolution tool, like something very specialized, it can often be more steps to get where you need to go, but you can get much closer to your final result.”
His solution: use Figma as the low-resolution base, then augment it with bespoke plugins that are high-resolution and specific to his exact use case. Each plugin might only be useful for four illustrations. That’s fine. The cost to build it is low enough that it’s worth it anyway.
The tool that makes the tool
Here’s what he’s actually doing.
He asks AI to build tools that let him make better illustrations than he could make by hand. The AI handles the precision work: calculating exact Gaussian distributions, running color quantization algorithms, generating accurate 3D kernel visualizations. Dan handles the design judgment: which examples to show, how to frame the visual, what the reader actually needs to understand.
As he put it: “Instead of just asking AI, can you make this illustration? You work out ways to help it help you.”
That’s the version of AI use I keep coming back to. The AI doesn’t replace the thing that makes your work yours. It removes the ceiling on what you can execute.
The oiler
Near the end of the video, Dan tells a story about Bell Labs in the early days of the telephone network. They had teletype machines running everywhere, and the standard oiling can made a mess. So a group of engineers built a very specific oiler that dispensed exactly 15 drops of oil to a precisely targeted spout, useless for any other application, but perfect for that one job.
Dan connects this to the plugins he’s been building:
“They’re not like useful to anyone else, but because the cost of building them is so low and I know enough about the problem space to make them, it’s just so satisfying to build these bespoke tools for yourself that do exactly what you need.”
That’s the frame I’ve been missing.
I’ve been building Andy, my AI executive assistant, for about eight months now. Friends and colleagues ask if I’m planning to turn it into a product. And the honest answer is: mostly no. Andy knows how I work, how I communicate, what I care about, and what I’m likely to forget. A lot of that is irreducibly specific to me.
The value was never in the tool being transferable. It’s in the fact that the cost to build something bespoke has dropped so dramatically that you can afford to build the oiler.
You can build a tool that does one thing exactly right for your situation, use it for the four projects where it matters, and move on. That wasn’t possible before. It is now.
What can I now make that I couldn’t before?
The question most people ask about AI is: what can I automate?
A more interesting question: what can I now make that I couldn’t make before, because I can afford the precision work?
Dan couldn’t have built those plugins without AI. The computation was too tedious, the iteration too slow, the margin for error too high. Now he can build them in an afternoon, use them for a chapter, and move on.
That’s augmentation. That’s the version worth paying attention to.
Speed is the easy part. The interesting part is capability: what you can now make accurately that was previously out of reach.