How to use Codex to speed up a puzzle campaign
Codex profiled Yodoku’s generator, caught a large-board optimization bug and enforced campaign pacing; Greco’s play feedback helped finish the Android game.
devpost.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Profile our deterministic puzzle generator in a separate worktree. Measure rejected candidates before optimizing, add a cheap no-guess screen, and preserve the existing campaign prefix with parity tests across every board size. Enforce visual-variety and recovery-level rules during selection, and stop before writing any invalid campaign.
Reach people exploring AI agents
See placements →What happened
The generator produced the reported 9,999-puzzle campaign while preserving deterministic parity during optimization; the Android game shipped, with iOS and cloud-save qualification still pending.
Reported by Austin Greco — DevpostHow the agent did it
- Codex measured generator behavior in an isolated worktree and identified discarded candidates as the main expense.
- It optimized screening and incremental updates, then replaced incorrect masks after comparison tests failed.
- Greco’s visual feedback became campaign-selection rules, and generation checks blocked an invalid pacing result before writing files.
Where a person stepped in
Greco judged puzzle variety, gestures, sound and native behavior; comparison and pacing checks caught agent mistakes before campaign output was accepted.
Austin Greco asked Codex to investigate Yodoku’s slow puzzle generator in a separate worktree. Profiling showed that expensive logical analyses were mostly being spent on candidates that stalled and needed guessing. The generator needed to preserve a deterministic campaign while rejecting unsuitable puzzles more cheaply.
Codex added an earlier no-guess check and updated only affected cells instead of repeatedly scanning the whole board. Its first optimization used integer masks that could not safely represent the larger boards. Comparison tests caught that before content was written, and the implementation changed to size-independent arrays.
The reported benchmark became faster while reproducing the existing puzzle prefix exactly. Greco then noticed visually repetitive early levels and asked for pacing changes. A final check stopped one full regeneration because a challenge lacked a recovery level. Moving that rule into selection produced the complete verified campaign.
The published Android game contains 9,999 puzzles. Greco made gameplay and sensory decisions and tested native navigation on a real device; Codex also repaired a predictive-back transition. iOS release and cloud-save qualification remained unfinished. The account shows measured optimization, parity protection and human play feedback working together.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Devpost · Austin Greco
Frequently asked questions
Why did the first optimization fail?
Integer masks did not safely cover every cell on larger boards.
Why check the old campaign prefix?
A speedup should not silently change deterministic output.
Was every platform complete?
Android was published; iOS and remaining cloud-save qualification were future work.
More details
About this account
Codex measured generator behavior in an isolated worktree and identified discarded candidates as the main expense. It optimized screening and incremental updates, then replaced incorrect masks after comparison tests failed.
Opened original dated creator account. Benchmarks and puzzle count are author-reported results. The source discloses a pre-event prototype and failed optimization/regeneration attempts; linked PR and log could not be fetched during this check. Suggested prompt written by Agents In Use; no original task prompt was found in the verified text.
Credit: Austin Greco — Devpost
Source published 21 Jul 2026
Source checked 14 Oct 2026
Tools and topics
Codex · Flutter · Flame · Dart · Kotlin · Swift
puzzle generation · parity tests · campaign pacing
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.