How the OpenAI team uses ChatGPT Sites daily
Kat, a website product leader at OpenAI, demonstrates how the ChatGPT Sites feature with plugin connectors enables diverse use cases ranging from incident management and team collaboration to creative projects like music discovery and 3D game development. The discussion highlights how Sites serves as both a practical business tool and an infrastructure platform for rapid prototyping and creative expression.
Summary
This transcript features a conversation between two OpenAI team members discussing the practical and creative applications of ChatGPT Sites, a feature that launched approximately two months prior to this Developer Day presentation. Kat showcases an incident command overview tool that allows product managers to monitor team progress on incidents in real-time by connecting to Slack, Notion, and Calendar with role-based data filtering—enabling each team member to see only data relevant to their team membership. The tool displays incident timelines, assigns responsibility, and links to runbooks without requiring complex API key configurations.
Beyond business applications, Kat demonstrates a music discovery site that aggregates songs listened to by her family members and surfaces popular Reddit playlists. She describes automating a weekly Monday morning playlist generation using Spotify as a connector, allowing real-time music discovery integrated into her work interface.
The conversation then pivots to more creative uses of Sites. One participant describes creating the largest user-generated dungeon exploration game, where players can build custom rooms using a 'Dungeon Site' skill that leverages Claude's Astra model to generate environments—from moon surfaces to volcanoes to Stranger Things-themed rooms. They discuss distributing skills rather than downloadable plugins, enabling others to create and contribute game rooms.
Another participant shares a real-time sketching app where drawing on canvas triggers Claude to generate corresponding imagery—if you draw a hill, a snail appears; if you draw a face, a hot air balloon appears. They emphasize how faster model inference (8x speed improvement) enables prolonged creative immersion and flow state, though noting it comes at 6x cost.
The discussion covers Sites as infrastructure, explaining that it handles deployment and builds integration with Claude, eliminating manual DevOps work. They also announce support for hosting MCP plugins through Sites. The conversation addresses analytics limitations, with the speaker preferring organic community discovery through Twitter hashtags and mentions rather than built-in metrics.
Finally, they discuss the 'awesomesites' community gallery showcasing user-created sites including bus route maps, piano applications, simulations, and Minecraft clones. The speakers emphasize how Sites democratizes software creation and enables people to express creativity previously constrained by technical barriers.
Key Insights
- Sites enables role-based data filtering where different team members viewing the same shared site see different data based on their team membership, allowing product managers to grant access without exposing cross-team sensitive information
- Faster model inference times directly enable sustained creative flow by reducing interruptions during the creation process, whereas slower model responses break immersion and suppress creative ideation
- The Codex model can recognize user intent sufficiently to suggest appropriate connectors and integrations proactively—for example, if a user requests 'a website for team collaboration,' the model asks whether plugin connectors should be used
- Sites handles all deployment and infrastructure concerns automatically when integrated with Claude, eliminating the manual DevOps work that users previously had to perform locally after code generation
- Distributing reusable 'skills' as game-building tools enables a model where players can directly request new content ('Please make a bunk bed') that Claude generates in real-time, rather than requiring developers to pre-create all game variations
Topics
Transcript
[0:00] We launched the site plugin feature today . With this feature, anyone can visit your site and use the connected plugins . In other words, you can only view your own data on that site. If I share this feature with my team members, will they see different data on their respective sites? Everyone views the same site, but the content is displayed differently depending on each team's role . It connects like magic without the need for complex settings like API keys. For example, there are times when you want to build something using ChatGPT or Codex and [0:30] think you need the Notion API or Slack API . It's really fun figuring out what these magical APIs…
Full transcript available for MurmurCast members
Sign Up to AccessMore from How I AI
I use Stripe Radar to catch card testing before it costs me
The speaker explains how to use Stripe Radar to prevent card testing fraud by leveraging built-in fraud rules, fingerprint detection, and Spot ID verification. These tools help identify suspicious payment patterns and validate that requests originate from legitimate app users, while allowing tuning to minimize false positives for good users.
Token fraudsters relayed their AI product through ChatPRD
The founder of Chat Purity discusses token fraud, where bad actors authenticate as their service and relay requests from their own AI products through Chat Purity to appear legitimate. Despite existing technical controls, fraudsters employed reverse proxies and relay systems to obscure their activity, creating inefficient and suspicious infrastructure.
He wants AI that keeps students doing the cognitive effort in math
The speaker emphasizes that AI in mathematics education should support student agency and cognitive effort rather than providing direct answers. Effective AI tools should interpret student ideas, encourage deeper thinking, and engage students in productive struggle while keeping them as active decision-makers in their learning process.
He put the family to-dos on a wall display Claude built
The speaker uses Dakboard, a wall display product, to make personal to-do items and information visible to their family. They had Claude build a script to scrape to-dos, upcoming trips, and other data from their database to display on the wall, including sections for immediate tasks, longer-term needs, and recent wins.
He built Claude Code a memory so he could trust its advice
A developer built a persistent memory system for Claude Code using markdown files stored locally on their computer, enabling the AI agent to maintain context across multiple sessions and provide better-grounded advice based on accumulated knowledge about the user.