How to use Codex to stabilize personal-best racing
Codex helped separate race and scenery randomness for reliable personal-best ghosts, then checked mobile behavior and local narration in a static browser game.
devpost.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Repair this personal-best racing comparison by separating visual randomness from deterministic simulation seeds. Preserve fixed-step physics, add repeatable contract-run tests and sector comparisons, and verify mobile profiles. Assemble offline narration from bounded clauses, then ask the human to judge race feel and remaining audio limitations.
Reach people exploring AI agents
See placements →What happened
The author reports a complete static racing game with deterministic personal-best comparisons and local compositional narration, with prosody and further device QA still pending.
Reported by Rhet Wike — DevpostHow the agent did it
- The team identified shared scenery and simulation randomness as the source of diverging personal-best comparisons.
- Codex helped separate seeds, add deterministic test seams and exercise desktop and mobile behavior through Playwright.
- The game combined local narration clauses and adaptive graphics, with human creative and acceptance decisions.
Where a person stepped in
Wike retained creative direction, gameplay feel and final acceptance; smoother narration and another physical-device QA pass remained planned.
Rhet Wike used Codex as an engineering and QA partner for a static orbital racing game. One specific defect undermined its personal-best rival: scenery and racing logic shared randomness, so nominally identical runs could diverge.
The project separated simulation randomness from visual randomness. A fixed 120 Hz simulation and comparable difficulty-and-contract seeds then supported a deterministic ghost with sector comparisons. Codex edited the implementation, created test seams, built Playwright regressions and analyzed screenshots as the team iterated on gameplay and mobile behavior.
Narration had a separate constraint. Complete prerecorded sentences could not cover every leader, overtake or finishing margin, while browser speech was inconsistent. A predecoded audio sprite assembled stock-voice clauses locally, avoiding a cloud request during the race. The author acknowledges that phrase-level prosody still needs improvement.
Human direction determined the concept, feel, visual judgment and acceptance. Adaptive graphics profiles reduced expensive effects on mobile while retaining a readable game. The reported result is a complete static game with deterministic comparisons and compositional race calls. It is not a claim that the fictional lab competition predicts real AI performance, or that every physical device received exhaustive testing. Another physical-device QA pass remained planned.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Devpost · Rhet Wike
Frequently asked questions
Why did identical races initially diverge?
Scenery and racing logic shared randomness; separating their random streams restored comparable simulation runs.
Does narration require a cloud request during play?
No. Predecoded stock-voice clauses are composed locally.
Was all device testing considered finished?
No. The author planned another physical-device QA pass.
More details
About this account
The team identified shared scenery and simulation randomness as the source of diverging personal-best comparisons. Codex helped separate seeds, add deterministic test seams and exercise desktop and mobile behavior through Playwright.
Firsthand dated Devpost account opened. Codex engineering and QA role explicit; gameplay narration and asset creation have separate tools. Fictional lab racing is not a real-model benchmark. Suggested prompt written by Agents In Use; no original task prompt was found in the verified text.
Credit: Rhet Wike — Devpost
Source published 20 Jul 2026
Source checked 14 Oct 2026
Tools and topics
Codex · Three.js · Playwright · AudioWorklet · Kokoro-82M
deterministic ghost · race randomness · offline narration
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.