A Good PM Feels the Pea: Don't Lower Your Standards When Building with AI

When building products with AI, I think it helps to have a little of the Princess and the Pea in you. A single pea under a thick stack of mattresses can still keep her awake. Products can feel like that too: something already works, but using it makes you uncomfortable. I think that feeling is valuable.

As I mentioned in a class, once a version of a page appears, a list starts forming in my head. Shouldn’t the avatar be visible here? Should the group avatar combine its members’ pictures? This part feels awkward; that part feels wrong. I call it my list of things that feel wrong.

Next, I prioritize those issues and address them one at a time, gradually building a complete version. You can’t work out every requirement at the beginning. Many problems only appear when you use the product yourself.

But some problems don’t go away, however much you adjust the existing approach. That’s when I step back: What am I trying to accomplish? Where am I now, and where do I want to end up? Once A and B are clear, the route between them can change.

A course is one example. Mention creating a course, and people immediately think of an outline, slides, recorded videos, and a platform to host them. But the outcome I want is for someone who couldn’t do something before to become able to do it—and actually use that ability.

Does getting there have to mean watching one lesson after another? I proposed an alternative: Could a course become a game with levels? Each level would involve a practical problem. The learner would need to apply the knowledge to complete the task. If they got stuck, they could receive a hint and try again. That shifts the question toward what we want them to accomplish.

This is a course design idea to try, with results that still need to be tested. Don’t reject it as soon as it appears just because courses have usually been made another way.

Of course, I research existing approaches first. If there’s a good solution, I won’t reinvent it. But if it doesn’t meet my requirements, I don’t lower them simply because everyone else works that way. I keep asking what doesn’t fit and whether another approach could work.

After years in an industry, you know what a product feels like when it works well. Your sensitivity to awkward details comes from that experience. You may once have accepted that software could only work a certain way. With AI, you can at least take that discomfort seriously and look for alternatives.

If the pea is still there, adding more mattresses won’t help me sleep.

EP0097 · 2026-09-13 · 02:44. The picture uses a digital human, and the voice uses my cloned voice. This is an English translation of the approved Chinese script used to create the narration.

When building products with AI, I think you need a little of the Princess and the Pea in you. There’s one pea beneath all those thick mattresses, and it still bothers her so much she can’t sleep. Building products is similar. Something clearly works already, but when you use it yourself, it just feels wrong. I think that feeling is incredibly valuable.

I mentioned in a class that once a version of a page appears, a list immediately starts forming in my head. Shouldn’t the avatar show up here? Should the group avatar be made from the members’ pictures? This part feels wrong, and so does that part. I gave it a name: my list of things that feel wrong. What happens next? I prioritize those issues and solve them one by one, gradually iterating toward a complete version. Could I have figured out all the requirements at the start? No. A lot of problems only appear when I use it myself.

But there’s another level to this. Some things still feel wrong no matter how much you adjust the original approach. That’s when I step back and ask: What am I actually trying to do? Where am I now, and where do I want to end up? This is what I mean by going from point A to point B. Get those two points clear first. The route in between can change.

I gave another example at the time: Could we turn a course into a game with levels? As soon as people hear “make a course,” a plan may pop into their heads: write an outline, create slides, record videos, and find a platform to put them on. The more experience you have, the faster that plan appears, right? But what outcome do I want? Someone who couldn’t do something before can now do it, and can actually put it to use. That’s the point B I’m trying to reach.

To get from not knowing how to knowing how, do they really have to watch one lesson after another? Could they work through challenges instead? For example, each level could be a real problem to solve. They’d have to use that knowledge to complete the task. If they got stuck, we could give them a hint and let them try again. See how the question changes? We start asking what we want them to accomplish, instead of what content we want to put into the course. As for whether that approach works well, we can try it and test it. Don’t reject an idea the moment it appears.

That’s what I meant in that sharing session. Often, before we even get started, the solution and its limitations have already appeared together in our minds. A course has to mean recorded videos; training has to mean explaining everything before letting people practice. Once you treat those as prerequisites, your conversations with AI stay on the same few familiar routes. It isn’t always that AI can’t do the job.

Of course, I’m not saying we should tear everything down and start over. I’ll definitely do the research first and look for best practices. If there’s already a good approach, I won’t reinvent the wheel. But if that approach doesn’t meet my requirements, I won’t lower them simply because that’s how everyone does it. I’ll keep thinking about what doesn’t fit and whether we can try another way.

So if you’ve spent years in your industry, don’t underestimate the things you find awkward or uncomfortable to use. You know what something should feel like when it really works well. That’s experience. In the past, you may have thought: Never mind, that’s just how software works. Now that we have AI, you can at least bring that discomfort into the open and seriously look for other approaches. If the pea is still there, adding more mattresses won’t help me sleep.

That’s it for today. See you tomorrow. Bye!