Last week I had to present my AI systems at a conference. I'd built the slide deck in Claude Design, and it was the best-looking presentation I'd ever made. The colors were arranged in beautiful gradients, the fonts looked professionally designed, and the flowcharts were animated. I was proud of it. But then, two days before the talk, I tried to make a few small edits and everything fell apart.
It’s easy enough to make a presentation in Claude Design. You chat with it or leave comments, just like you work with any other AI tool. But under pressure, I couldn’t make changes fast enough.
In Claude Design, you don't drag a box the way you would in PowerPoint. You describe the change in words and let the tool rebuild the slide. When I asked it to delete two of the slides, the comments I'd left on the other slides came loose, and it lost track of what belonged where. The fix would have taken ten seconds by hand, and instead it was taking twenty to thirty minutes flibbertigibbeting between responses.
I tried exporting the slides instead, and it came back wrecked: the formatting was a mess, the fonts were wrong, and the gradients I had loved flattened into ugly pink squares. With the clock running, I panicked.
So I went back to the tool I already knew. Google Slides is plainer and it can't animate your flowcharts, but it let me move a box myself. I stayed up until 2am the night before my presentation, rebuilding the deck by hand.
My failures that day pointed at a lesson, and it's bigger than slides. AI gets you eighty percent of the way in thirty seconds. The deck that looks almost done, the draft that reads mostly fine, the email that's basically there. But you don't live in the first draft. You live in the edits: the swap, the cut, the reorder, the fix at eleven at night.
I've written before (in Issue 10) that the hard twenty percent is the judgment. The part where you decide what's good. But this was a different twenty percent. I knew exactly what I wanted: delete two slides, move a box, add a new photo. But in the moment, communicating those changes to the AI tool was the part I couldn’t crack. The hard part here wasn't deciding what to change. It was fighting with the AI to change it.
The fix is obvious in hindsight: before you trust a new tool with something important, spend some time learning the tool, and trying to break it. Build something small, then put it through the edits you'd need in a crunch, while you’re not yet in one:
Can you nudge a box “a smidge to the right”, or do you have to just hope it works out?
When you edit the copy, does it change in place, or does the whole piece rebuild?
If you add, delete, or move something, does the rest of it hold together or do parts break or disappear? (Every time I duplicated a slide, images kept disappearing from the deck!)
When you export a file, or move to another tool, how much of it survives intact?
How long does an actual edit take? Ten seconds? Or ten minutes of back-and-forth?
Spending fifteen minutes on questions like those would have told me everything I learned the night before my presentation back when I still had time to fix it.
The talk went well. The audience got what they came for, despite the lack of animated flowcharts. And luckily the time spent battling Claude Design wasn’t a loss. It was an impressive tool and the slides were better than anything I'd have made alone. Plus building them was fun. Next time I'll just have to learn how to edit slides before the deadline, not the night before.
The whole thing in 30 seconds
I built a beautiful slide deck in Claude Design, then couldn't make small edits under time pressure. I rebuilt it in Google Slides at 2am the night before.
AI gets you eighty percent there in seconds. The last twenty percent, the edits, is the hard work. And sometimes it’s difficult to communicate the changes to the tool, even when you know exactly what you want.
Judge a new tool by how easily you can change what it makes, not by the first draft it shows you.
Spend fifteen minutes editing in it and trying to break it before you trust it with something important.
Don't meet a tool's limits for the first time on a deadline.
What's a tool you loved but never made the time to learn? Hit reply, I read every one.
me+machine