What GPT-5 builds when nobody touches the keyboard

What it does

The repository houses a curated set of front-end demos—websites, games, and interactive interfaces—that OpenAI says were each created by a single natural-language prompt to GPT-5 or GPT-5.2, with no hand-written code. Visitors can inspect the zero-shot prompts behind each example and treat the collection as a reference for what the model can scaffold from description alone. A hosted version lives on Vercel, and the repo includes a local front-end viewer.

The interesting bit

This is less an open-source project and more an official certification wall: OpenAI curating its own model’s one-shot output to set baseline expectations for hands-free generation. The repo doubles as a product funnel, steering developers toward the Codex CLI and non-developers toward ChatGPT’s Canvas preview.

Key highlights

• Every demo was generated in a single prompt session with GPT-5 or GPT-5.2, according to the README.
• Each example exposes the exact zero-shot prompt used to create it, making the gallery a prompt-engineering reference.
• The collection is strictly front-end scaffolding: websites, games, and interactive UIs built from natural-language descriptions.
• OpenAI labels the repo as reference-only and is not accepting contributions.
• It explicitly points users toward Codex CLI for terminal workflows and ChatGPT Canvas for browser-based preview.

Caveats

• The demos are a curated selection chosen to showcase the model’s strengths, so expect a highlight reel rather than a representative sample.
• Scope is narrow to front-end toys and UIs; there are no backend or systems-level examples.
• Contributions are closed, making the repo a read-only gallery unless you fork it for your own experiments.

Verdict

Worth a browse if you are evaluating GPT-5 for rapid front-end prototyping or looking for effective one-shot prompts to adapt. Skip it if you need backend patterns, production-grade architecture, or an open-source project to contribute to.