ISSUE 30 · ME+MACHINE
Sometimes I give AI a sentence and it produces something that feels like magic. This time, I explained exactly what I wanted. I provided context, examples, detailed instructions, and the benefit of my excellent judgment…
And it was an aesthetic emergency.
Over the last few days, I’ve been trying to turn ideas from the Gimbal Podcast into social media carousels. There’s useful information in those conversations and I wanted to share those ideas beyond the episode itself.
The early results were surprisingly good! This was my first attempt.
Not bad, right? A few small edits, I thought, and I could post these.
Naturally, I was already fantasizing about how I’d set this up to run automatically after every episode. I captured the process in an AI skill—a reusable set of instructions for making the next carousel.
I worked on it all day, adjusting the copy, trying visual directions, and working through what I liked and disliked. I built up a small collection of decent one-off carousels.
But by the end of the day, I was disappointed. The automated results kept trending back towards a generic AI-generated look.
The oversized headlines, scribbles, arrows, and scraps of paper were all competing for attention. The original slides had distinct images that helped tell the story, whereas these repeated the same decorations around different text. They were louder but less memorable.
Each correction seemed to take the results further from what I wanted.
Luckily, I recognized this pattern from working with AI on my writing. “Make it more conversational” or “cut the salesy language” sounds like useful feedback, but doesn’t tell the AI what a good revision would look like. I had started keeping voice files and writing guidelines to give the AI something more concrete to work from.
I needed similar guidelines for visuals, but I’d never written them before. I could tell when a carousel looked bad, but explaining why, or what would make it better, was harder.
That made it difficult to judge the instructions the AI wrote, too. I hadn’t examined them closely enough, and I didn’t always know whether its descriptions matched what I wanted. Some of its own suggestions had ended up in the skill as preferences I’d never expressed.
I don’t know how much those instructions contributed to the disappointing carousels, but before adding more, I needed to check what I was already asking it to do.
Decide what’s worth repeating
By the end of the day, our conversation contained experiments, corrections, abandoned directions, and a few results I liked. Before turning that into instructions for the next episode, I should have separated out which parts were worth keeping.
That would have meant looking at what each instruction was supposed to achieve. If I asked the AI to remove an arrow because a slide was too busy, saving “don’t use arrows” could just lead to another decoration taking its place. A more useful instruction would explain that visual elements should help communicate the idea, rather than simply fill space.
Something that worked on one slide might not suit another. I should have made clear which instructions applied to every carousel and which depended on the content.
I also needed to remove the AI’s own suggestions that had become instructions I’d never asked for.
Show it what I want instead
For the next version, I need to explain what a good result looks like. Visual examples could help me show what I mean when I don’t have the design vocabulary to describe it.
I can start with examples from the first carousel and explain why I liked them. The strongest slides used a clear visual idea to communicate the point. The image would change with the content, but its job would stay the same. That gives the AI a direction to follow instead of leaving it to replace one decoration with another.
If I struggle to explain a preference, I could put a slide I liked beside one I don’t and ask the AI to describe the differences. I’d use that description to find the words, then check that they expressed what I meant before adding them to the skill.
Test the skill on its own
I’ve already tested the skill on other episodes in fresh conversations, with the material and references it was supposed to use. Once I’ve revised the instructions and added examples, I’ll run that test again.
When a result disappoints me, a favorite trick is to ask the AI to compare it with the instructions. I’m looking for two different problems—did it ignore a clear requirement, or did it follow a rule that needs changing? Its answer gives me a place to investigate, though I still check it myself.
What I'm taking from this
I still think I can automate this workflow. But I underestimated how much I’d need to learn to give it useful direction. I could recognize a carousel I liked long before I could explain what made it work.
That’s the part I need to get better at. With writing, I’ve built up language and examples to explain my preferences. I’m only beginning to do that for visuals. AI can help me put those preferences into words, but I need to decide whether it’s understood me before saving them as instructions.
THE 30-SECOND VERSION
A promising one-off doesn’t prove that a workflow is repeatable. Test the skill in a fresh AI chat with new material.
Keep instructions that explain what you want and why. Separate rules that apply every time from choices made for one slide, and remove AI suggestions you never approved.
Use examples to show what good looks like, and explain which qualities should carry over. Check the AI’s interpretation before saving it as guidance.
When a result disappoints, compare it with the instructions. Check whether the AI missed a requirement or followed a rule that needs changing.
This time, my excellent judgment will be reviewing the instructions.
me+machine.
