CueChaos Inside Saboteur

Inspiration

AI party games often become either a solitary prompt box or a stream of unrelated random jokes. We wanted a social game where friends debate real choices, remember consequences, and catch a human bluff without giving strangers a way to drain the creator's API budget.

What it does

CueChaos gives 3-6 players one connected mission with a deadline and concrete stakes. One player secretly becomes the Inside Saboteur. Across three causal phases, the group debates three plausible plans and votes anonymously. The selected plan changes case progress and risk, reveals a concrete clue, and becomes visible history in the next phase.

The Saboteur knows one tempting high-risk plan per phase and tries to make at least two win while protecting a case-specific motive. After three phases, everyone privately accuses a suspect. The investigation team wins only by identifying the Saboteur and preserving the mission.

The game ships with four complete cases and supports English and Chinese. It is designed for one shared phone and takes about 8-12 minutes.

How we built it

GPT-5.6 worked directly inside Codex as a controlled case writers' room. It created and refined the four bilingual missions, roles, motives, decisions, consequences, clues, and cause-and-effect chains. We used Codex to audit every case with a simple standard: if the phases could be reordered without breaking the story, the story was not ready.

Codex was also the primary environment for gameplay design, implementation, responsive styling, tests, the zero-API security model, deployment, video preparation, and submission packaging.

The app uses TypeScript, React, Next.js, Vinext, Vite, and a Cloudflare-compatible Workers runtime. A local route validates 3-6 unique names, deterministically deals one Saboteur, rotates roles and plan order, and serves the reviewed case pack. Ballots, case state, accusation scoring, and the verdict all run locally.

Challenges

The hardest design problem was making every option plausible while keeping a real causal story. An obviously bad option makes the Saboteur trivial to spot. An unrelated option makes the story feel random. We iterated until each choice had a rational pitch, an immediate consequence, evidence, progress gained, and risk created.

The second challenge was controlling cost and abuse. The deployed game contains no OpenAI credentials and makes no runtime inference calls. GPT-5.6 is used during the controlled Codex build workflow, then the reviewed case packs are committed to the repository. Visitors cannot trigger billable inference.

Accomplishments

We are proud of the complete pass-the-phone experience: private roles, anonymous ballots, persistent evidence, final accusations, deterministic replay, responsive bilingual UI, four coherent cases, a public credential-free demo, and automated tests for schema integrity, determinism, validation, causal plans, and zero external API calls.

What we learned

Generative AI works best here as an accelerated creative and engineering partner, not as an uncontrolled runtime dependency. Codex helped us move between narrative design, implementation, testing, visual polish, deployment, and documentation while keeping every generated case reviewable.

What's next

Next we would add community-authored case packs with a validation tool, accessibility options for larger groups, optional local sound cues, and a case editor that uses the same causal constraints before exporting a safe static pack.

Built With

• codex