Build one mechanic. Then hit Play.
The whole-game prompt fails for a
boring reason. When everything is
generated at once, you cannot tell
which step broke the loop. Summer
Engine’s 70% ceiling is the useful
map: the model can draft movement,
wire a state machine, stub UI, and
drop placeholders. Feel, timing,
signature art, and the one track
people hum still sit on a human seat.
Unity can sell a casual game from a
sentence. GDC’s floor is still using
the same models to argue with scope
and write the email.
A Godot afternoon can copy that split
without buying a new engine. Ask the
chat model to name the smallest
version that is still fun, and the
five systems you are underestimating.
Do not let it agree that the
roguelike-deckbuilder-farming sim is
a weekend. Then give the coding
assistant one job: a dash, a petal
pickup, a store-page screenshot
scene. Run it. If coyote time is
wrong, you know which node to open.
Generated sprites can fill the item
row. The hero pose, the HUD, and the
jump still get a human pass.
Stop collecting tools. One assistant
that can see the scene, one chat
model for scope and copy, metered art
and audio only when the placeholder
is not good enough for Play. If
“make it juicier” starts adding
systems, uninstall the clever layer
and ship the mechanic you can already
feel.
Checklist
-
Ask the model to cut scope, not to
bless it. Smallest fun version, five
underestimated systems.
-
Prompt one mechanic. Hit Play.
Name what broke before you add the
next node.
-
Keep generated art in the bulk
(icons, tiles, beds). Human-pass
the identity pieces: hero, jump,
store screenshot.
-
Treat generated 3D as a kitbash
base. Retopo, UVs, and a rig before
it is a character.
-
Draft barks offline. Runtime NPCs
are a live cost and a different
design.
-
If you cannot point at the one
thing that got better in Play,
delete the prompt and keep the
handmade loop.
This afternoon: one mechanic, one
Play pass, one store-blurb draft.
No second system until the first one
feels like a game.