Apex Driving Simulator — a Gran Turismo-grade racing sim in one HTML file, where a harsh critic sub-agent measured the track spline until the asphalt stopped lying
A single-file three.js racing simulator built by fanning out sub-agents against a Gran Turismo 7 benchmark — Pacejka-style tyre slip with relaxation-length transients, ABS/TC/ESC, a 6-speed box with real gear ratios, procedural asphalt, kerbing, telemetry traces and a ghost lap. Then a separate critic agent tore it apart in a real headless browser: it sampled the road's RGB (128,130,135 at 0.054 saturation) to prove the tarmac was neutral grey and not green, hammered the barriers at 161 mph to prove 0.0000 m penetration, and measured the centreline spline point-by-point to find a 7.28-degree kink at the loop closure that was silently causing three separate visual bugs.
Follow along on video
There is a walkthrough for this example on the Learn with Kody channel. It takes the build end to end in eight numbered steps — the exact seed prompt below, what the agent actually returned, the moment the obvious fix failed, the measurement that found the real cause, and the re-run that proved it. Pause at any step and do the same thing.
Watch the walkthrough (2:40) · Drive it live, with no video file · Subscribe to Learn with Kody
The channel is published on RAPP Vision, a local-first video platform served entirely as static files from this account. The second link is not a video at all: it is a few KB of JSON that loads the real simulator and replays a recorded script on a clock, so you can pause it and take the wheel.
What this is
A single self-contained HTML file that is a full racing simulator. No build step, no npm, no asset pipeline — three.js from a CDN and everything else generated procedurally at load: the asphalt, the grass, the kerbing, the trees, the sky.
The physics is the point. Each tyre runs a Pacejka-style combined-slip model with a relaxation-length transient, so grip builds and decays over distance travelled rather than snapping instantaneously. Weight transfers longitudinally under braking and laterally through corners, changing the normal load on each corner and therefore the available grip. A 6-speed gearbox with real ratios sits behind a torque curve. ABS, traction control and stability control are individually toggleable and each one measurably changes your lap time. Telemetry traces throttle, brake, steering and slip along the bottom of the screen, and a ghost of your best lap runs alongside you.
But the reason this demo exists is the loop that built it. A builder sub-agent wrote the simulator. A completely separate critic agent — running its own headless Chromium over http, not file:// — then tried to destroy it, and was forbidden from accepting any claim that wasn't a number it measured itself.
That critic found things no static check could. It sampled the road surface and got RGB [128,130,135] at 0.054 saturation against grass at [48,82,41] and 0.499 — proving the tarmac reads as neutral tarmac. It drove head-on into a barrier at 161 mph and asserted maximum penetration of 0.0000 m. It computed a 12-bit colour histogram of the viewport from outside the page to prove the chase camera never clips inside scenery, driving the worst dominant-colour frame fraction from 0.50 down to 0.111.
Why this is mind-blowing
The best bug in this build was one that looked like a texture problem and was actually a geometry problem.
By the start/finish line, the asphalt showed a fan of about fifteen dark radial streaks. It read unmistakably as broken skid-mark decals. The builder agent agreed, cleared the skid buffer, reported skidActive() === 0, and declared it fixed. The fan was still there.
So the critic stopped looking at pixels and walked the track's 1600 centreline points directly. Median per-point heading change: 0.378 degrees. Maximum: 7.28 degrees — nineteen times the median — and the six worst values were at indices 1594, 1595, 1596, 1597, 1598 and 0. That is the seam where the loop closes. At the same place, segment length collapsed to 0.194 m against a 0.743 m median.
The spline was not C1-continuous where it closed on itself, and it was being sampled by curve parameter rather than arc length, so points bunched into a cusp at the seam. That one defect produced three symptoms that looked completely unrelated: the asphalt texture's transverse seams compressed into a radiating fan (the "skid marks"); the road ribbon's inner edge offset exceeded the local radius of curvature and folded back over itself, leaving a jagged triangular notch in the grass boundary; and Reset spawned the car at index 0, inside the sharpest 40-degree bend on the circuit, which is why holding the throttle with no steering pinned you against a barrier in three seconds.
Fixed by reshaping the closure control points, resampling by arc length with arcLengthDivisions of 6000, and deriving tangents from centred finite differences on the periodic positions. After: max per-point turn 2.12 degrees, segments 0.648 / 0.649 / 0.650 m, and the worst curvature relocated to index 727 — an actual corner. Holding throttle from Reset now stays on track through 63.7 mph.
Three bugs, one root cause, found by refusing to look at the screenshot and reading the numbers instead.
Prompt
I want you to build a driving simulator at the level of Gran Turismo 7 and
Assetto Corsa. It should be utterly perfect, visually beautiful, with every
single thing done at AAA quality — from the tyre model to the asphalt
textures to the suspension geometry to anything you could think of.
Real Pacejka-style combined slip. Weight transfer under braking and
cornering. A proper 6-speed gearbox with real ratios and a torque curve.
ABS, traction control and stability control that each measurably change lap
time. Telemetry traces for throttle, brake, steering and slip. A ghost lap.
Fan out sub-agents and have sub-agents tackle each one individually so that
the sim is utterly perfect. You should /loop on each item and have a
separate sub-agent check it VISUALLY in a real browser to ensure it looks
and behaves triple A. That separate sub-agent should be a really harsh
critic, and if it doesn't look triple A, it should keep going.
The critic is not allowed to accept prose. Every claim must be a number it
measured itself: sample the road pixels and prove the tarmac is neutral grey
and not green. Drive into a barrier at speed and prove the penetration depth
is zero. Measure the centreline spline point by point and prove there is no
discontinuity. If the critic cannot produce the number, the fix is rejected
and goes back to the builder.
Don't stop until each sub-agent is utterly wowed with the quality when
compared with the actual Gran Turismo. It should literally compare them side
by side blind and say which one looks better. Do this in ThreeJS as ONE
self-contained HTML file. /loop until it's utterly perfect. Fan out
sub-agents and ultracode.Paste this into Claude, Cursor, or Copilot. Change one thing that matters to you.
What I learned shipping it
- How to make a sub-agent critic useful by banning prose: require a measured number for every claim, so 'it's fixed' becomes 'penetration is 0.0000 m at 161 mph' and unverifiable success reports stop being accepted
- How to find a spline discontinuity by walking the centreline and comparing per-point heading change against the median — a 19x outlier at the loop-closure seam explained three unrelated-looking visual bugs at once
- How to stabilise a tyre model at v->0 where the slip ratio (omega*R - v)/|v| goes singular, using a relaxation-length transient plus a low-speed force fade instead of clamping