Cases / 01 · 3D city kart

Should the kart
use the skill?

The short answer: yes, if the race is the product. No, if the race is a physics test. One brief, two loops, same model, same single-file constraint. The no-skill build is already a complete game. The skill build is a more finished one.

Verdict: use itDeepSeek V4.1 FlashA/B harness comparisonNot a model benchmark
Use itwhen the deliverable is a finished race

The verdict in one paragraph

The control build, Sundown Circuit, is 890 KB and already fun: three laps, six karts, a city loop, drift and nitro. It also has a thin shell. The skill build, Sunset District, is 1.58 MB and adds settings, save, touch, an objective chain, five rivals, a four-lap structure, a wet-road reflection system and a title screen that reads as key art.

That is a product difference, not a rendering difference. If you are shipping the race to other people, the skill earns its rounds. If you are testing whether a kart can drift, skip it: the control got there without the gates.

01The same brief

Two loops,
one prompt.

Both runs used DeepSeek V4.1 Flash, a single self-contained HTML file and no external images, models or audio. The only intended variable was the harness: the procedure plus its gates.

Sundown Circuit: a kart race through a city street, with HUD, minimap and boost pad

A / no skill

Sundown Circuit

3 laps · 860 m · 18 turns · 6 karts · 890 KB

Sunset District: karts on a wet downtown street at golden hour

B / with the skill

Sunset District

4 laps · 0.49 km · 5 rivals · 636 draw calls · 1.58 MB

DimensionA / Sundown CircuitB / Sunset District
Harnessnoneskill + eight gates
File size889,961 B1,577,115 B
Race3 laps · 860 m · 18 turns4 laps · 0.49 km · 10 turns
Rivals5 AI karts5 AI karts, named, with catch-up
ShellTitle, pause, win/lose, touchTitle, settings, pause, save, win/lose, touch, quality presets
Signature surfaceWet-look road, no planar reflectionWet asphalt + half-res planar reflection + puddle mask
WorldCity loop, props, crowd100+ city blocks, 2 landmarks, 20 prop families, 435 meshes / 759 instances
Acceptance—8/8 gates · product, mobile, census, lights, playtest, look

Numbers from the shipped files and the project gate output.DeepSeek V4.1 Flash / 2026-11

02What the skill changed

The lap was already there.
The product wasn't.

The control proves the model can already build a racing loop. The skill did not teach it to drift. It made the shell, the content budget and the surface system non-optional.

01 / SHELL

Six screens, one save file.

The skill build treats the title as key art: HUD hidden, premise, start, settings, controls and best run. Settings persist. Touch is reachable at 390×844. The control has a title and a pause; the skill build has a product.

product gate · mobile gate

02 / CONTENT

An objective chain, not a lap counter.

Four laps, five rivals, cone hazards on the racing line, boost pads, damage and repair, a live classification and a podium condition. The loop has a shape: catch up, hold the line, survive the clock.

content budget

03 / SIGNATURE SURFACE

The road is a system.

Wet asphalt is not one roughness value. It is an albedo/roughness/normal set, gutter and tyre-rut wetness, separate puddles, clearcoat and a half-resolution planar reflection injected into the road material — quality-gated, not free.

look + performance budgets

04 / GATES

What turned red.

Product caught the missing shell. Mobile caught touch reachability. Census caught low-segment geometry at close range. The stock collision gate assumes a walkable rectangle, so the racing contract was re-expressed as barrier push, wall clamp and track containment.

census · collide · playtest

03The cost

It is not free,
and it is not automatic.

The skill build is 77% larger, took more rounds and added systems that the control did not need. It is a more finished product, not necessarily a more fun prototype.

Sunset District: a kart drifting through a wet corner, with the race HUD
Signature surface: wet road, reflection, puddlesSunset District / shipped build

The skill buys the last 20% that makes it a product.

If the race is already fun at 890 KB, the extra 687 KB is not a rendering upgrade. It is a shell, a content budget, a surface system and a gate log. That is worth it when someone else will download the game; it is waste when you are testing a drift model.

The control also has something the skill build can lose: speed of iteration. A smaller file with fewer systems is easier to change. The skill's product-first order protects the player's first minute, but it does not make the core loop better.

04From draft to product

The title changed
because the shell did.

The early build had the race. The final build had a title screen that could sell it. The difference is not a filter; it is hierarchy, a premise line, a start action and a saved best.

An early Sundown Circuit title screen

EARLY / functional

The race exists. The first frame does not yet argue for it.

The final Sundown Circuit title screen

FINAL / key art

Title, premise, start, controls and best run over the live scene.

05The gate log

What "finished"
was allowed to mean.

The gate list is the argument. The racing-specific collision and playtest contracts were adapted for a continuous barrier corridor; the stock walkable-rectangle assumptions do not transfer.

shadersNo compile/link errorPASS
census435 meshes / 759 instances, 0 failuresPASS
collideBarrier push, wall clamp, track containment (adapted)PASS
lights4 lights, all anchoredPASS
playtest31/31 racing checks (adapted)PASS
productShell, content, save, touch, budgetsPASS
mobile390×844, touch, no overflowPASS
lookTone/emptiness fallback; flat 0% heroPASS
measured / sunset district
draw calls  636
triangles   539,883
lights      4
meshes      435
instances   759
size        1,577,115 B
network     0 requests

# the output that mattered
product    PASS
mobile     PASS
census     PASS
lights     PASS
playtest   PASS  (31/31 adapted)
look       PASS  (fallback)

The look gate used the declared fallback because the project deliberately did not use an external photographic reference. That is stated, not hidden.

06Verdict

Use it when,
skip it when.

The same rule applies to any genre: the skill protects the player's first minute and the reviewer's ability to check the build. It does not make the core verb fun.

USE IT

You are shipping the race.

Other people will download it, play it on a phone, change settings and come back. The title has to work. The save has to persist. The gates earn their keep.

product · mobile · save

SKIP IT

You are testing a drift model.

One track, one kart, no shell, no save. The control build got there without the gates and is easier to change. Do not pay for a product layer you are not shipping.

prototype · physics · look test

ADAPT IT

The collision contract is genre-specific.

A racing circuit is a continuous barrier corridor, not a walkable rectangle. Re-express the contract on the page: barrier push, wall clamp, containment and progress. The gate should run on an unrelated subject; the page's test plan changes.

collide · playtest

Play both sides

The control and the product are one click apart.

Read next

The shooter case.

Saltline is the reference. Neon Breach is the no-skill control. Ironsight is the green-but-thin first skill run. Black Ice is the second generation: 318 KB, 8/8, four roles and a boss.

Read the shooter case