(I Forget What They're Called)
The Setup
I volunteer as a coach for my boys’ hockey league. I’ve been building a small AI project that helps me turn practice plans into short animated clips.
I was doing some research on how to shore up our defensive zone and still apply the American Development Model (ADM) approach. So I stretched my practice plan tool a bit to draw our defensive zone, which is called “the house,” and to show where the kids should skate and why.
I put a lot of care into the brief, especially the language. One line I wrote: “Kids should never have to ask “where?”“ Every place on screen gets one name, and it’s the same name the coach uses on the ice.
My Approach + AI Role
I built it with Claude Code in one sitting, maybe 30 to 45 minutes. Before building anything, Claude flagged about 11 places where my brief conflicted with the project’s own rules. I answered quickly and a little carelessly. At one point I typed “not sure what you’er saying but sure, try it out.” 😬
Then it built the clip, and I watched it on my phone.
What Actually Happened
The first version was good enough to react to. Then I noticed the camera zoom was stuttering.
I told Claude. It found a bug, fixed it, and the stutter was still there.
On the second try, I had a hunch in the back of my head. So I asked Claude if it was a refresh rate problem.
Claude picked that up and ran with it. It explained that the clip was playing at 30 frames per second, which can look choppy on a phone, and moved it up to 60. That made sense to me. It was good reasoning, and it was partly right. The motion did get smoother.
It was still stuttering.
By the third round I was frustrated enough to write, “I need you to be thoughtful about this and not take the path of least resistance.” Claude built a tool to check the motion frame by frame. The tool said everything was smooth. It wasn’t.
On the fourth round, I described exactly when the stutter started and how it got faster as the zoom sped up. That’s what finally led to the real cause. Claude had been checking the motion it intended, not the frames it actually drew. Once it checked the real frames, the problem showed up right at my timestamps. After the fix: “much better.”
Then I asked Claude what we’d missed, and I typed this:
“I used to edit videos often. I’ve used After Effects. I know how to plot those (I forget what they’re called)…”
Keyframes.
I’ve used them for years. I just couldn’t remember the word. Once I explained what I wanted the way I’d do it in After Effects, everything got easier. Claude had been building a video tool from scratch, and the person who knew how video editing works was me. I just hadn’t said so. 🤦
The Real Insight
I assumed Claude could understand what I was picturing. But I was picturing something After Effects would have made, and I never mentioned After Effects.
The funny part is that my whole brief was about making sure kids never have to ask “where?”
Then I spent the session learning to tell an AI the things I didn’t think to say.
I think this is what working with AI keeps teaching me. It reflects back what we already know. We know more than we think, but we forget, and we take it for granted. The wrong outputs are what help us notice it.
It’s a bit like getting to know a neighbor. You can have a lot in common and still make wrong assumptions about each other. And you can’t always be perfectly clear up front, because you don’t know the other person well enough yet to know what needs saying.
So you guess, and sometimes you guess wrong. That’s okay. That’s normal. Some of my best friendships started with a misunderstanding we ended up laughing about. That moment of friction is often what moves a conversation past small talk and into something deeper.
Working with AI has felt similar. I couldn’t be clear about something I didn’t realize I knew. I had to see the wrong version first, and the misunderstanding is what got us to a better understanding.
That’s the difference I keep feeling between “AI made this” and “I made something with AI.” The clip only got good once I brought what I know to the table.
Try This If…
You’re working with AI on something you’ve done before in another tool or craft:
- Tell it how you’d do it there. “In After Effects, I’d do it this way” would have saved me a few rounds.
- Build a rough version early, so you can see what’s wrong.
- When the AI says it’s fixed, ask it to show you.
Systems Lens
Most of the talk about AI at work starts with efficiency. Do more, faster, with fewer people. Lately I hear more companies soften that. They say AI is about being more effective, or freeing up time for more meaningful work. But both versions still assume the same thing: that the skill is already in the AI, and we just hand it the task.
This week showed me where that falls short. What I needed was in my head. Part of it was a hard skill, knowing how a tool like After Effects works. But it came bundled with judgment: where a pause should sit, how a zoom should ease in, when something feels choppy even if I can’t say why. Those are tiny decisions I make without thinking. I’ve made them so many times that I stopped noticing I was deciding anything.
The AI had plenty of skill. It could build a renderer, check frames, and reason about refresh rates. What it didn’t have were my small decisions. And I couldn’t hand those over to the AI until I remembered I had them.
I think that changes the question for anyone bringing AI into their work. It’s easy to wonder which of our skills AI will replace. This week, an old skill was exactly what I needed. I’d just stopped noticing I had it.
So the more useful question for me now is: what do I already know that I haven’t said yet?