Kath Korevec’s ChatGPT Sites Workflows: Dashboards, Music, and Games

Live from OpenAI DevDay, I had the chance to sit down with someone I immediately recognized as a "sites twin"—Kath Korevec, the product lead for ChatGPT Sites at OpenAI. We both share a passion for pushing the boundaries of what you can build on the web with AI, and Kath's work is a perfect example of this. While many people might think of Sites as a place to host a simple webpage or a prototype, Kath showed me how her team uses them to create truly dynamic, personalized software.

In this episode, we moved from the purely practical to the purely fun. Kath walked me through three distinct workflows that showcase the versatility of ChatGPT Sites. First, we looked at a serious work tool: an incident command dashboard that pulls in live, user-specific data from tools like Slack and Notion. Then, we tackled a common problem for any parent—rescuing a music algorithm from the clutches of children's songs—with a custom music discovery app. Finally, we explored the creative frontier with a collaborative 3D dungeon crawler game that the community can build upon.

These aren't just toy examples; they represent a fundamental shift in how we can think about building applications. By combining ChatGPT Sites with the power of connectors, custom skills, and models like Codex and Astra, you can create tools that are perfectly tailored to how you and your team work and play. I'm Claire Vo, and this is How I AI.

Workflow 1: An Incident Command Dashboard with Live Data

As a product manager, Kath is often juggling meetings and high-level strategy, but she still needs to stay informed about technical incidents without getting in her team's way. Constantly pinging engineers for updates isn't efficient for anyone. The solution? She built a custom Incident Command dashboard using ChatGPT Sites, which gives her a real-time overview of what’s happening.

What makes this so powerful is the use of connectors, a feature available for enterprise workspaces that brings secure, authenticated data into your Site.

The Step-by-Step Process

Here’s how this personalized dashboard comes together:

1. Connecting the Data Sources: The Site is hooked up to the team's core tools using connectors for Slack, Notion, and Google Calendar.
2. User-Specific Authentication: When a team member visits the Site, it asks them to approve these connections. The Site intelligently uses their credentials to pull data that is specific to them.
3. Contextual Awareness: The dashboard is smart enough to identify which team you're on. By looking at the Slack channels a user has access to (like #sites-private or #sites-feedback), it infers they are on the Sites team and customizes the view accordingly.
4. Displaying Live Information: The dashboard presents a timeline of current incidents. Kath can see what's happening down to the minute, who is on-call, and what the status is. She can even click to open the relevant runbooks and guides directly in Notion.

The beauty of this is its scalability. If someone from the Codex team visited the same Site URL, they would see their team's incident data, not Kath's. This eliminates the need to build separate dashboards for every team and ensures data remains properly scoped and secure.

The Magic Incantation

How do you tell the model to build something like this? It's about expressing intent. You can describe the goal plainly: "Build me a website that can handle bringing in Notion documents and cross-reference that across my calendar." Kath also shared a magic word she uses:

"My magic word is like 'bring your own connector.' or 'Can you use plugins insights to build this site?' that kind of thing."

This directs the model to leverage the connector architecture, creating an application that feels deeply integrated with a user's personal and team data.

Workflow 2: A Personalized Weekly Music Playlist

Like many of us with kids, Kath's Spotify algorithm was... compromised. Taken over by Bluey and Sesame Street, it was no longer a reliable source for discovering new music. To solve this, she built a simple, mobile-friendly web app called "Heavy Rotation" on ChatGPT Sites.

This workflow is a fantastic example of building personal software—a tool that solves a unique problem for you and is tailored to how you want to use it.

How It Works

The goal was to create an automated system for music discovery that required zero weekly effort.

1. Sourcing New Music: Kath wanted to tap into what's genuinely popular, so she instructed the Site to find music recommendations from Reddit. She uses the computer use capability, which allows the model to browse websites.
2. Automating the Process: She set up an automation to run the discovery process every Monday at 8 a.m. The model scrapes Reddit for top playlists and popular tracks that people are sharing and upvoting.
3. Building and Playing the Playlist: The Site then compiles a new playlist for the week. The interface itself is simple and uses the Spotify web player to stream the music, but it can also reference her local music files on Apple Music.

The result is a fresh playlist every Monday morning, filled with music she’s never heard before, completely separate from her main Spotify account. It’s a perfect little application that reclaimed her music discovery experience.

Workflow 3: Building a Community-Driven Dungeon Crawler Game

This is where things get really creative. Kath and I discovered we had both been experimenting with building 3D games with the latest models, and her project highlights a fascinating approach to community collaboration. She built a 3D dungeon crawler and, more importantly, created a system for others to add to it.

The project isn't just about building a game; it's about creating a framework for shared creation.

Part 1: The Base Game

Kath started by building the initial game on a ChatGPT Site:

• A 3D dungeon crawler set in a digital version of San Francisco's Alamo Square.
• Players can choose a character and a "Codex pet" to join them.
• The gameplay involves exploring the world and battling enemies (in this case, teddy bears).

Part 2: The 'Dungeon Site' Skill for Expansion

The most interesting part is how others can contribute. Instead of a complex level editor, Kath created a custom dungeon site skill.

1. Defining the Rules: This skill is a set of instructions for Codex that defines what a "room" is—its dimensions, player controls, object properties, etc.
2. Creating a New Room: To build a new part of the dungeon, a user installs the dungeon site skill. They can then use a natural language prompt with a model like Astra to describe the room they envision.
3. Natural Language Creation: For example, this morning Kath's son was watching the Stranger Things cartoon, so she created a new dungeon room with that theme.

She just described the theme, and the model, guided by the dungeon site skill, generated the entire room. You could ask for anything:

"I want to be on the moon surface."

"I want to be in a volcano."

1. Connecting the World: The skill even asks the creator if the room should be "open-ended" (allowing it to connect to other rooms) or a final level.

Part 3: The Community Vision

Kath's idea is to use this system to build the "largest user-created dungeon crawler." Creators build their rooms, share them via Twitter with a specific hashtag, and over time, these rooms can be linked together into one massive, sprawling game world. It’s a brilliant way to use skills as a distribution model for user-generated content.

Awesome Sites Gallery

To see what else is possible, Kath pointed me to Awesome Sites, a gallery she built to showcase cool projects from the community. It's inspired by the classic "awesome lists" and is a testament to people's creativity. You can check it out at awesome-sites.ai.

We saw everything from a fully functional piano built with Sites to a detailed, interactive bus atlas for Hyderabad, India. This gallery is a perfect place to get inspired and see how people are using Sites to build games, utilities, and art.

Conclusion

My conversation with Kath Korevec was a fantastic look into the future of personal and collaborative software. We saw three distinct applications of ChatGPT Sites: a highly practical work dashboard, a fun personal utility, and an ambitious creative game. Each one demonstrates that Sites are far more than a simple hosting platform; they are a surface for deploying dynamic applications powered by AI.

The common thread is the move away from rigid, one-size-fits-all software. By leveraging connectors for live data, using skills to enable complex creation, and building on fast, powerful models like Codex, we can now build tools that adapt to our individual needs. Whether you're a product manager needing a specific view of your team's work, a parent wanting a better music playlist, or a creative who wants to build worlds with others, the tools are now accessible enough to make it happen.

I hope this inspires you to think about what "personal software" you could build. What unique problem can you solve for yourself or your team? Go check out the gallery at awesome-sites.ai for inspiration, and start building.