Vibecoding
Build for the vibe, debug later
Start new thread

The screen that demos best is the one our daily users skip fastest

by•

Arteza opens on a grid showing all six studios at once, image, video, voice, music, avatars, editing. It demos brilliantly. It's the screen people screenshot and it's why anyone understands the product in four seconds.

It's also the screen our heaviest users leave immediately. If you're in it every day you already know which studio you want, so a menu of six is six things to skip. The people who love that grid are the ones seeing it for the first time, and they only see it once.

The mistake underneath is that I designed the first thirty seconds and assumed the next thirty days would sort themselves out. Everything I measure at signup rewards that grid. Nothing I measure rewards the tenth session, so the tenth session never got designed.

Trying to make the home screen a function of what you did last week instead of a fixed grid, which is easy to say and quietly a different product. If you shipped something that demoed well, what did you have to change once people actually lived in it?

48 views

Add a comment

Replies

Best

Same split on the marketing side, except I ship pages, not screens. The shot that sells is the empty state with everything laid out, and the shot that makes existing users nod is the messy one with 200 files in it. I ended up putting both on the page, the clean one up top and the ugly one further down where people are already sold.

 Putting both on the page is smarter than what I did. I only shipped the clean one, and the thing new users say in week two is that it doesn't look like what they saw. Did the messy shot move anything you can measure, or does it just stop the complaint later?

  Mostly the second thing. I can't separate it from everything else changing on the page, so I won't claim a number, but the complaint you're describing stopped showing up. My read is it filters rather than converts, and a page that filters costs you some signups you were going to lose anyway.

The 'going to lose them anyway' part is the bit I'd want to be sure of. With us those people don't leave in week one, they leave in week three, so the cost lands somewhere I'm not measuring. That said, filtering is still the right trade here. Our grid converts a stranger in four seconds and then sets an expectation the tenth session can't meet, and I'd rather lose someone at the screenshot than at the invoice.

  Fair, and your week three is the part I can never see. The page hands someone off and I never find out which of them quit later, so I only ever get the version of the story that happens before signup. Losing them at the screenshot is the one place it's still my problem, which is probably why I like that trade.

  You can see it if you tag it. We stamp the signup with which page variant sent them and then don't look until day 21, and the traffic that converts fastest is usually the traffic that's gone by then. It's one URL param and one column, and it took about an hour to wire up. I'll send you our week three split once the new home screen has enough people through it to compare.

  The URL param is the half I can add myself. The column is a favour I'd have to go ask for, which is probably why I've never had one. Yes to the week three split whenever it's ready.

  You don't need the column. We put the variant on a PostHog person property before we ever touched our own database, and that answered the week three question for about four months. Asking someone for a migration is how this quietly never happens. I'll ping you with the split.

Asad naming the screen his heavy users skip is the harder half, and I fell into the same shape. I published ten convictions from two and a half years of building this way, and the eighth says a demo wins attention while only repeated results earn trust.

Then I had to write at the bottom of the same piece that my production test is five runs and the library is one company's. Conviction eight cuts at everything above it.

@cmumulle Five runs and one company's library is the honest version, most people publishing ten convictions leave the sample size out entirely. Where I'd push is that conviction eight is a measurement problem before it's a belief. Repeated results only earn trust if someone sticks around for the repetition, and every event I instrumented fires in the first session, so I had nothing that could tell me whether run ten went better than run one. I designed the demo and then measured the demo.