Same-Prompt AI Game Build Comparison
Two agent-built runner games are compared through split-screen gameplay, escalating map reveals, and live presenter reactions.
Starter prompt / by Refs
The creator shared no short prompt. This is ours for making a film like it with Claude.
Code a colorful 3D endless runner set in a New York subway, as one HTML file with three.js. The hero rides a glowing hoverboard down three lanes of track, switching lanes, jumping barriers and ducking under signal gantries as trains roll in, and grabs spinning gold coins. Give it chunky stylized art, a station platform with tiled pillars and stairs, a score and coin HUD, and a steadily rising speed. Record a 60-second autoplay run at 1920×1080 with headless Chrome and ffmpeg.
Field notes / why it works
The first frame states the shared test above simultaneous moving outputs, and the first cut waits until 0:07.9 so viewers can compare for themselves. A's result gets a specific 0:12–0:32 critique; B then reveals three increasingly distinct areas at 0:32, 0:54 and 1:04 before the 1:27 split-screen verdict. The 6.52-second speech gap around 1:04 lets the strongest visual reveal carry the argument.
Plate 01 / The format in one breath
Give two tools the same creative task, show their outputs simultaneously, then leave the split screen to inspect each result in turn. The promise is immediate visual evidence of a fair comparison; the second result keeps revealing new features and the presenter's opinion intensifies with it.
Why it works
- The cover and first frame label both sides and say “SAME PROMPT · SAME DAY”; the test is intelligible before the first cut at 0:07.9.
- The first result gets a compact critique at 0:12–0:32, while the second gets three escalating locations from 0:32 to 1:27.
- The glowing city section at 1:04–1:22 turns the verdict into a visual discovery; the presenter pauses for 6.52 seconds after 1:04.4 while gameplay supplies the reveal.
Plate 02 / Format card
| Platform · aspect · length | X · 16:9 · 1280×720 · 60 fps · 97.4 s |
| Pace | 25 detected shots, 24 cuts; 14.8–19.7 cuts/min allowing uncertain changes; 17.3 definite picture changes/min; mean shot 3.9 s, median 3.07 s |
| Script | 228 measured words; 146 wpm across the runtime, 181 wpm while speaking; 19 sentences averaging 12 words |
| Voice | One animated, conversational medium-register presenter, initially evaluative, then increasingly surprised |
| On screen | ~12% side-by-side comparison, ~15% first gameplay alone, ~55% second gameplay with small reaction inset, ~10% second gameplay without prominent inset, ~8% presenter close-up (estimated from sheets) |
| Captions | All-caps white chunk subtitles near bottom centre; heavy black outline; selective blue/yellow emphasis and fixed game labels |
| Sound | Continuous quiet tonal/game-music-like bed, speech forward; no clear cut-synced effects; −14.0 LUFS integrated, −0.9 dBFS true peak |
| Structure | Same-test hook → first result → second result/location escalation → surprised reaction → split-screen verdict; no explicit CTA |
Pro / 13 plates + recipe
The first plates are open. The full dissection is Pro.
Get the timed shots, type and sound specs, and the runnable recipe.
Open the full blueprint- Plate 03 / The hook, frame by frame (first 3 seconds)Pro
- Plate 04 / Structure (beat sheet)Pro
- Plate 05 / Shot grammarPro
- Plate 06 / Visual style guidePro
- Plate 07 / Voice and deliveryPro
- Plate 08 / Sound designPro
- Plate 09 / PackagingPro
- Plate 10 / Script templatePro
- Plate 11 / Production planPro
- Plate 12 / PromptsPro
- Plate 13 / Edit recipePro
- Plate 14 / QA checklistPro
- Plate 15 / Make it yoursPro
- Recipe / runnable promptPro