Author the generator. Then hit Play.
Two different weekends hide in the same prompt.
One weekend writes a function: given a seed, build
a cave that connects, lights, and loads in a
frame. The other weekend asks a model to dream
the next room while someone is already holding
the stick. The first weekend ends with a file
you can check in. The second weekend ends with
a demo you cannot reproduce.
Classic procgen is still the product. Noise,
adjacency rules, room stitches. A language model
is a fast intern for that code. It is a bad
runtime. Latency, cost, and drift are not vibe
problems. They are why shipped games still roll
dice the old way. Keep the model in the author
seat. Keep the tick boring.
The store does not care which seat you used.
Play’s API floor moved under Godot 3 this week.
An editor assistant that paints a prettier icon
will not get an AAB through review. Check the
export target, bump it, install on a phone, then
generate. The play is the review. Everything
before the play is inventory.
Checklist
-
Open the Android / iOS export preset first.
Read the target SDK and the min OS. Write them
down.
-
If Play wants API 36 and you are still on 34,
bump and install before you generate anything
new.
-
Use a model to write or tune a deterministic
generator (noise, WFC, BSP). Do not call a
model to invent rooms at runtime.
-
Same seed must reload the same layout. If it
does not, it is not a level yet.
-
In-editor assistants stay on drafts — lines,
tags, variations — until the result is in a
running scene on a device.
-
If prompting plus import took longer than
placing three rooms by hand, retire the tool
for this job.
This afternoon: confirm the export floor, then
land one seeded generator (or one handmade room)
that plays the same twice. That is the weekend.