Nobody Talks About the Waiting

Every AI feature I've built has the same shape.

You press the button. Then you wait.

Ninety seconds for a blueprint. A couple of minutes for a render. Long enough that you can't do anything useful, short enough that you don't go start something else. So you sit there and watch a bar fill.

Nobody puts that in the demo video.

---

I've shipped a lot of generation features into PodGlue this year, and the honest truth is that the waiting is part of the product whether I designed it or not. If I don't decide what happens in that gap, the answer is that you leave the app and go check something else. Then you come back four minutes later, or you don't.

So this week I put a game in there.

It's called Untangle. It's in the Games tab in the iOS app. Coloured dots in pairs, and you draw a line between each pair without crossing another line. Ten levels, ten puzzles each.

That's the whole thing.

---

Here's the part that's actually interesting, and it's not the game.

The puzzles are generated, not hand made. A hundred hand made boards is a week of work and I would have gotten some of them wrong. So the app builds them from scratch, the same hundred boards for everyone, and a test proves every single one is solvable before it ships.

My first approach was the obvious one. Pick two dots that are far apart, then find a route between them. Do that for every pair.

It produced zero valid puzzles. Not a few bad ones. Zero, out of a hundred.

The reason is a thing I should have seen and didn't: if two pairs straddle each other, there is no way to connect both without crossing. Think of the four corners of a square. Connect the top left to the bottom right, and you have just cut the square in half. The other two corners can never reach each other. It doesn't matter how clever you are with the line.

So most spread out dot layouts have no solution at all. I was generating puzzles that were impossible and then acting surprised when the router failed.

The fix was to turn it around. Grow the line first, wandering across the board, then put the dots at the two ends. Now the route is the proof. If I drew it, it can be drawn.

---

The second thing I got wrong was harder to see, because it looked fine.

I wanted the puzzle to make you go around something. That's what makes it feel like a puzzle instead of a chore: the direct line is blocked, so you have to think.

I wrote the code, and then I measured how many pairs were actually blocked. Almost none.

Turns out a dot sitting in the way blocks nothing. On a grid there are a dozen equally short ways to get from A to B, so a single obstacle just gets stepped around at no cost. The puzzle looked harder and wasn't.

There's exactly one case where a single dot does block you, and that's when both dots of a pair sit on the same row. Then the straight run is the only shortest way, and one dot parked in the middle forces you the long way around.

So now the game builds that on purpose. Level one has one of them. Level ten has four.

I only found that because I measured instead of trusting that it looked right. That's the lesson I keep relearning.

---

I want to be straight about what this is. It's a puzzle game in a podcast tool. It doesn't help you book a guest or ship an episode.

It's there for the ninety seconds.

But the ninety seconds are real, and pretending they aren't is how software ends up feeling like a chore. The waiting was always part of using PodGlue. Now at least something is happening in it.

If you're on the iOS app, it's in the Games tab. Start on level one. The dots that look easy to connect are the ones I put there to get in your way.