I built a word game recently. WordSlip. It took a fraction of the time it would have taken five years ago. The animations are smooth, the UI is clean, it looks and feels like a proper game
It’s not a good game.
The core mechanic doesn’t work. In a good word game like Wordle, there’s a natural process of discovery. You guess, you learn something, that learning guides your next guess. Every move gives you a small hit of progress. That’s what makes it satisfying.
WordSlip has a discovery mechanic too. But to find a letter in a valid position, you typically have to construct a word that isn’t valid. That’s not a natural thing to do while playing. So instead of a loop of guess, learn, progress, you end up staring at your available letters until you’ve mentally brute-forced the answer. Then you put it in. That’s not a game. That’s a chore.
I shipped it anyway. 25+ years making games and I didn’t catch it.
Friends told me. They couldn’t guess the word because they never naturally did the discovery step. In hindsight it’s obvious. Historically I would have found it myself, because I would have spent weeks with the prototype, feeling the mechanic, letting my gut tell me if the loop worked.
This time I didn’t. AI made the surface-level quality so cheap that the game felt finished before I’d properly played it. The polish came so fast that I skipped the step that actually mattered. I got blinded by how easy it was to make something that looked good. And I neglected the thing that built my career: obsessing over whether the core mechanic is any good.
WordSlip is what I’d call a false positive. A product that passes every surface-level check but fails the one that matters.
AI is good at producing false positives. Ask it to build something and you’ll get a competent version. Sometimes impressive. But “works” and “good” are not the same, and the gap between them is where products die.
When building was expensive, this gap was smaller. The cost of production forced you to spend time with what you were making. Weeks of building a prototype are also weeks of playing it, living with it. You’d feel what was wrong before you could articulate it. That’s not mystical. It’s exposure.
AI compresses the building so much that the living-with-it part disappears. I went from idea to polished product so fast that my evaluation process never kicked in. And I’ve caught myself doing this outside of game development too. Review something, looks fine, move on. The work is flowing, stopping to evaluate kills the momentum. So you let things through.
The danger isn’t that AI produces bad work. It’s that AI produces work fast enough to outrun your judgment.
I’d bet there are more polished-looking failed products right now than at any point in history. Not because founders got dumber. Because the thing that used to slow you down long enough to notice problems has been removed.
The lesson from WordSlip isn’t “be smarter” or “have better taste.” I have 25+ years of domain expertise and I still shipped a broken game. The lesson is that my process failed, and I need to fix the process.
Here’s what I’ve changed:
Prototype before polish. I would have caught this if I’d put a bare prototype in front of five people before touching animations or visual design. I didn’t, because it already felt finished. The polish made the game feel done when the mechanic wasn’t. Now the rule is: ugly prototype first, user reactions second, polish last.
Make evaluation explicit. When development was slow, evaluation happened naturally. You’d spend so many hours with the thing that problems surfaced on their own. That doesn’t happen anymore. If evaluation isn’t a deliberate, scheduled step in your process, it won’t happen. AI removed the forcing function, so you have to build a new one.
Don’t let generation speed set review speed. AI can produce ten things in the time you used to produce one. That doesn’t mean you can properly evaluate ten things in that time. The generation got faster. The judgment didn’t. If you’re reviewing at the speed of generation, you’re rubber-stamping.
Domain expertise still matters. When my friends told me WordSlip wasn’t fun, I could immediately trace the problem to the discovery mechanic. Someone without that background would get the same feedback but might spend months tweaking the UI, the onboarding, the difficulty curve. Everything except the thing that’s actually broken.
But domain expertise alone doesn’t save you if your process lets you skip the evaluation step. I am the proof.
I’ll be honest about something: I’m not sure discernment stays a competitive advantage.
Right now, the combination of AI tools and genuine domain expertise is rare. That’s the window. But it’s narrowing. And there’s a harder question underneath: what if volume and measurement replace discernment entirely?
Ketchapp built a two-billion-download business by receiving 100 game submissions per week, publishing the ones that tested well, and using cross-promotion to scale winners. No deep understanding of why games worked. Just volume and measurement. In advertising, this has gone further: publishers like RiverGame now run 90%+ AI-generated creative, testing thousands of AI-generated variations per week, no human judgment in the loop, $20 million a month in revenue.
For ads, brute-force search has already won. For products, the loop is slower and the cost per experiment is higher. Building a game that retains users for months is harder than finding an ad that gets clicks for seconds. But every improvement in AI generation and automated measurement closes that gap.
When it does close, the advantage won’t be who can build the best product. It’ll be who can put it in front of the right users. Distribution survives after creation commoditizes. If you have audience now, you have an advantage then.
That’s what I’m doing. Using discernment while it still matters. Building audience while I still can.