Devlog: two Lemmy posts brought 23 first time players and none of them came back. What the data said, and what I changed.
Short version: the channel worked and the game did not retain a single person. Here is the autopsy, because the numbers turned out to be more specific than I expected.
The game is a shared 256x256 pixel canvas where an AI adversary erases what people paint. Free, browser, no account. Two posts, one in a pixel art community and one in a web games community, brought 150 arrivals, 38 sessions and 23 people who actually painted. That is more than the project had seen in its entire life. Of those 23, one came back to look and zero painted on a second day.
What the sessions looked like:
- 18 of 23 lasted two minutes or less. Median time from first pixel to last: 41 seconds.
- Median 19 pixels painted. 9 of 23 spent all 30 charges (the bucket refills one charge every 30 seconds).
- In 14 of 23, most of their cells were cells taken back from the adversary. They were not decorating, they were fighting her.
- In 18 of 23 she erased cells while the player was still there, and between 126 and 423 more in the 25 minutes after they left.
- Nobody used undo. Zero comments on either post.
What I read from that: the first session is 30 pixels long, it ends when the charges end, and nothing happens at that moment. No closure, no name, no reason to come back. The adversary was already reacting to them, but she is forbidden from touching any pixel younger than 72 hours, so her answer always landed somewhere they were not looking.
What shipped since, in two passes.
First, making her legible in the first minute:
- When your bucket hits zero she speaks to you, by name, once in your life. The zero stops being a bare number and becomes a state with a countdown.
- Her answer to the cells you took back travels signed, so you can tell it was aimed at you.
- The identity prompt stopped interrupting the burst.
- Her blitz finally respects the 72 hour protection on young pixels, which the code promised in writing and did not do. That single bug had eaten 238 of the 403 cells that cohort painted.
- Instrumentation for the whole first session, so the next wave is measured instead of guessed.
Then, a deeper problem: subtraction is invisible. Nobody notices what is no longer there. A tower sixteen storeys tall, which costs 24 charges of attention, vanished in one hit that cost her exactly what a flat pixel costs. So:
- Height is now resistance. One hit takes one storey off a tower; only a flat light goes out in a single blow. A first session is still as fragile as it was, because nobody builds storeys in their first minute, but what you invested in now holds.
- And she does not only erase, she converts: every storey she takes she raises as her own cold cube in the void cell beside it, and the ring of a siege now rises a storey, so you watch the cage close around a work before the blow lands. Her towers fall in one pixel when you reclaim the cell. Hers is brittle, yours is stubborn.
- And the welcome screen stopped being eight paragraphs nobody read. It is now a four step animation that shows how to build, what her hit does and how to take a cell back, with one line of text under it.
This post is part of the second wave. The success criterion was written down before publishing so I cannot move it afterwards: at least one in five of the people who reach zero charges spends a second batch, and at least 3 of the new painters paint again within 72 hours. Baseline is 0 and 0.
The thing itself, if you want to look: https://ultimaluz.com/r/lemmy-gamedev-lw
One critique I would genuinely like: is a hostile adversary the wrong shape for a first session, or is the shape fine and what was missing is an ending?
8 replies
Is this entire post AI generated, or just blocks of it?
The entire thing stinks of Claude
The "game" is you paint by pixels and a chatbot erases them...
The entire thing is likely AI slop, including all the social media spam
Fair question. Parts of it, yes. I'm a Spanish developer and English is not my first language, so I wrote the post with an AI assistant and edited it afterwards. The numbers, the decisions and the mistakes are mine. I also build a good part of the game with AI tools and I'm not going to pretend otherwise; the manifesto page on the site says who is behind it.
What is not AI is the canvas. Every lit pixel there was placed by a person, one click at a time. The AI in the game only erases, on a budget, with its rules published in the open.
If the writing put you off, that's on me, and it's useful to know.
There's some different opinions on the role of AI in game development. Some people only accept it as a coding assistant. Some people say no to even that. But I think you'll find that most people on Lemmy in general say they would rather interact with a person a little broken English than interact with AI against their will.
Please consider others when making a post for a community. Also consider that if people begin to think "oh, is the person that uses AI all the time" every time they see you post, they may avoid your posts and your work in the future. You don't want that either.
Seems to lag on my phone, instructions moved too fast, even after going back to read them I'm still not sure what the point is
Thanks, this is exactly the kind of thing I can't see from here. Which phone, if you don't mind? The 3D city is heavy on older Androids; the flat map (the PLAN button) is much lighter, and the glow can be switched off with the button next to it.
On the instructions: the four dots above the animation let you step through it at your own pace, but if you didn't notice them, that's my design failing, not you reading badly. The point in one line: light pixels, keep them lit, and an AI is trying to put them out.
I'm on a fold 7 I figured out how to step through the steps again fortunately, it was just a bit fast automatic. I learned the objective, but not the point... It might be better to say I learned the mechanics but didn't get hooked on the gameplay loop