In 30 seconds
Kath Korevec shows Claire Vo an incident dashboard, a music discovery site, games and an experimental widget. She works on Sites, and the discussion distinguishes a live demo problem from finished capabilities.
- 1Incident overviewThe incident overview combines team information in a demo.
- 2Connected dataVisitors can use their own connections subject to access.
- 3SourcesThe guest uses common work connectors for personalized views.
- 4InfrastructureSites can host software and, as described, MCP plugins.
- 5Music discoveryA music site combines personal instructions and outside sources.
- 6GamesReusable skills support collaborative game experiments.
- 7Experimental widgetThe inference widget is experimental and its live demo encountered a problem.
Takeaways
- The incident overview combines team information in a demo.
- Visitors can use their own connections subject to access.
- The guest uses common work connectors for personalized views.
- Sites can host software and, as described, MCP plugins.
- A music site combines personal instructions and outside sources.
- Reusable skills support collaborative game experiments.
- The inference widget is experimental and its live demo encountered a problem.
Read the edited text by chapter
Work and connected data
An incident overview and personal connections.
Claire Vo interviews OpenAI product lead Kath Korevec about using ChatGPT Sites for work, personal projects and games. Korevec works on the product, so the demonstration reflects an insider’s perspective; the episode has sponsors.
Korevec shows an incident command overview for her team and emphasizes that the displayed content is a demo. The site brings together information from connected work tools to help the team see incidents and related context.
The pair discuss using connected services as data sources. Korevec describes a feature in which people visiting a shared site can use their own connected plugins and see data according to their access. The exact behavior depends on the connector and permissions.
Her frequent sources include Slack, Notion, Google Drive and Calendar. She describes both small personal tools and team software, while noting that connector coverage continues to evolve.
Infrastructure and music
Deployment and personal discovery.
Korevec describes Sites as deployment infrastructure for software built with Codex, and says it can also host MCP plugins and data for other workflows. Feature availability and permissions depend on the user’s environment.
For a personal example, she describes creating a mobile-friendly music discovery site after family listening changed her recommendations. It gathers suggested music and helps build a playlist; the example combines external sources and user instructions.
Games and experimentation
Collaborative projects and a prototype widget.
The discussion turns to games. They describe building interactive projects and sharing them with others, including using reusable skills so people can contribute new material to a collaborative game.
Korevec shows an experimental widget intended to bring model inference into a site. The demo does not work as intended at that moment, and she describes the feature as something the team is exploring rather than a finished guarantee.
Community and boundaries
A gallery and control over external actions.
She shares a community gallery of examples, from a piano to a bus map, as a way to discover what others have built.
Asked about prompting, Korevec says she gets frustrated when an agent makes unwanted assumptions, particularly when it contacts people on her behalf. She explicitly tells it not to do that, highlighting the value of clear boundaries around external actions.
More from this channel
35:10EN
How I AI· Oct 28, 2026PodcastFour Personal AI Agents on Nine Tasks: A User’s TestHost Claire VoClaire Vo compares work and family workflows while acknowledging uneven access and subjective scoring.
29:40ENHow I AI· Oct 26, 2026Podcast · 2 speakersHow Reddit Proto Connects Design, Live Data and Production CodeGuest Blair Metcalf · Host Claire VoBlair Metcalf demonstrates Reddit’s internal prototyping harness and its design-system mapping for Figma and Cursor.
16:22ENHow I AI· Oct 21, 2026PodcastClaire Vo: Claude Duet Is Beautiful, but It Is Not My Number One AgentHost Claire VoAnthropic’s phone-based personal agent has the most polished launch Claire Vo has seen: thoughtful onboarding, task lists, first-class artifacts, and Slack-style channels. Polish is not enough when every other agent is competing for the same attention.
36:33ENHow I AI· Oct 19, 2026Podcast · 2 speakersHow Brooke Lacey Uses Claude Code to Edit Short VideosGuest Brooke Lacey · Host Claire VoA creator combines local media tools, timed cut lists and human feedback.
24:05ENHow I AI· Oct 14, 2026PodcastA First Look at Meta Muse Wonder on Agent and Design TasksHost Claire VoClaire Vo reports strengths and weaknesses from her own small evaluation.
37:47ENHow I AI· Oct 12, 2026Podcast · 2 speakersDesigning Meta Muse: Status, Personalization and Everyday AgentsGuest Alex Cornell · Host Claire VoAlex Cornell discusses consumer agent design and his own experiments with Muse.