Agents In Use
What’s new todayLast update
1005 additions

How to use Codex to generate solvable murder mysteries

Codex built Cold Case with backward crime generation, unique-culprit checks, and one seeded random stream, publishing a replayable offline mystery game.

OUR ADAPTATION

Suggested prompt

Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.

Build a single-file offline noir mystery from this design spec. Generate the culprit first, independently supported innocent alibis, and exactly the contradictions needed for one solution. Reject failed cases before play. Route every random choice through one seeded PRNG, then test shared seeds, clue accessibility, interrogation, evidence-board, timeline, and accusation behavior.

What happened

Cold Case was published as an offline single-file mystery game with shareable deterministic seeds and a generated-case solvability verifier.

Reported by Aayan Zahoor and Munhim Qazi — Devpost

How the agent did it

  1. The creators specified mystery mechanics, tone, and unique-solution rules for Codex.
  2. Codex generated crimes backward and rejected cases that failed the unique-culprit check.
  3. It kept randomness in one seeded stream and implemented interrogation, evidence, timeline, and accusation flows.

Where a person stepped in

The creators wrote the design specification and directed bugs, interface tweaks, and narrative polish through the same implementation session.

The Cold Case creators gave Codex a design specification for a replayable noir mystery. The hard requirement was fairness: each generated case had to expose enough evidence to identify one culprit, rather than leaving the player with an arbitrary hidden answer.

Codex implemented a generator that constructed the crime backward. It selected a killer, gave innocent suspects independently supported alibis, introduced the killer’s false account, and placed contradicting evidence. A solvability check rejected a generated case unless the culprit remained uniquely identifiable.

The same session built interrogation, a draggable evidence board, an updating timeline, and accusation outcomes. Every random choice used one seeded generator so a shared seed reproduced the same cast, clues, and dialogue. The creators iterated on bugs, interface changes, and noir presentation through that session.

The result was an offline single-file browser game. The authors report stress-testing generated cases across difficulty settings, but this account does not turn their finite checks into a proof about every possible seed. The useful pattern is pairing procedural variety with a verifier that can reject an unfair puzzle before the player receives it. Human direction still defined what counted as fair and what the story should feel like.

Prompt
Suggested prompt (written by Agents In Use, not the author's)

Source: Devpost · Aayan Zahoor and Munhim Qazi

Frequently asked questions

How did the generator preserve fairness?

It built the culprit and alibis first, then verified one uniquely contradictory suspect.

Why use a single seeded random stream?

The same shared seed must reproduce the same case and dialogue.

Does the game need a model while playing?

No. The published result runs offline as one HTML file.

More details

About this account

The creators specified mystery mechanics, tone, and unique-solution rules for Codex. Codex generated crimes backward and rejected cases that failed the unique-culprit check.

Original team account lists verified named creators alongside a private user. Design specification was described but not shared verbatim, so prompt is suggested. Finite author-reported stress tests are not universal solvability assurance. Suggested prompt written by Agents In Use; no original task prompt was published in the verified account.

Credit: Aayan Zahoor and Munhim Qazi — Devpost

Source published 21 Jul 2026

Source checked 14 Oct 2026

Tools and topics

Codex · HTML · JavaScript · seeded PRNG · browser

procedural puzzles · deterministic seeds · solvability

Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.