ChatGPT Sites at OpenAI: Internal Uses and 60 Connectors

A product manager at OpenAI explains how ChatGPT Sites is used internally, the accompanying tools, and an ecosystem of around 60 connectors. She also details concrete use cases, her prompting method, and the limits she sets for the AI acting on her behalf, with links to resources and public galleries.

Accessing Official Resources and the Public Gallery

Direct links are provided to ChatGPT Sites (chatgpt.com/sites) and OpenAI Codex (openai.com/codex), as well as to the OpenAI DevDay page (openai.com/devday) and the Awesome Sites gallery (awesomesites.ai). Contact information for Kath Korevec is included, notably her LinkedIn profile (kathleensimpson) and her X account (simpsoka). Links to ChatPRD (chatprd.ai), Claire Vo's website (clairevo.com), as well as her LinkedIn (clairevo) and X (clairevo) profiles are also listed.

Internal Uses, Connectors, and Underlying Infrastructure

Kath Korevec describes an incident management site designed for her team at OpenAI, which uses Slack and Notion connectors in real-time. The ChatGPT Sites connector ecosystem includes about 60 integrations. She presents her favorite connectors and original uses of the Sites. A single phrase allows Codex to configure connectors for the user. An internal infrastructure layer within the Sites, still little explored by users, is also mentioned. The presentation revisits internal testing and the product architecture.

Kath Korevec Discusses the Launch of Plugin Insights and Hosting MCP Plugins

Kath Korevec explains her involvement in the deployment of Plugin Insights and the hosting of MCP plugins. She details how Plugin Insights works and its impact on the public that can use a site. She reports having built and used ChatGPT Sites internally for over a year before the public launch, and specifies her role within the OpenAI product team working on Codex.

Examples, Creativity, and Limits of AI Action

Demonstrations include creating a personalized Spotify playlist and fixing a Spotify account using Reddit and computing tools. The design and development of games are also discussed, featuring a community dungeon crawler based on a skill distribution model, presented as a new approach to collaboration. The integration of inference into the Sites via a widget experience and the Awesome Sites Gallery are highlighted, along with Kath Korevec's prompting strategy. She asserts that the model's speed influences creativity and sets a clear limit on what the AI can do on her behalf.