CHAPTERS
- 0:00 – 1:32
Plugins in Sites: per-user data views without sharing credentials
Kath explains the newly launched “Plugins in Sites,” which lets anyone visiting the same Site use their own authenticated connectors. That means the Site experience can be shared broadly while each viewer only sees data they’re permitted to access—no API key wrangling for the Site creator.
- •Plugins in Sites enables visitors to use their own connected plugins/connectors
- •Same Site URL, but content varies by viewer’s team context and permissions
- •Designed for Enterprise workspaces with approval prompts for connections
- •Eliminates need to distribute API keys or manually manage auth
- •Enables Sites to function as secure, role-aware internal tools
- 1:32 – 4:58
How OpenAI uses Sites internally: incident command dashboards and team ops
Kath shows an incident command overview Site her team uses, rebuilt with demo content for safety. The Site pulls from tools like Slack, Notion, and calendar data to create a live, navigable view of ongoing incidents, people involved, and runbooks.
- •OpenAI has used Sites internally for over a year; keynote slides were a Site
- •Incident command Site aggregates incident timelines and status details
- •Pulls runbooks/guides from Notion and links directly into source docs
- •Uses Slack channel context to infer the team scope (e.g., Sites team)
- •Supports team workflows like OOO calendars and lightweight team sharing
- 4:58 – 5:54
Sponsor break: Merge (integration + permissions infrastructure for AI products)
Claire shares why integration, permissions, and production infrastructure become the hardest part of shipping AI products. Merge is positioned as an infrastructure layer that connects to thousands of tools and helps agents act securely while optimizing routing and spend.
- •AI products often get bogged down in integrations, permissions, and reliability
- •Merge connects to thousands of tools and provides secure agent actions
- •Positions itself as a production infrastructure layer for AI
- •Mentions customers like OpenAI, Dropbox, and Ramp
- •Call to action: merge.dev/howiAI
- 5:54 – 7:53
Prompting Codex to use connectors: “intent-first” building in Sites
Claire and Kath discuss how builders can ask Codex/Sites for apps that draw from Notion, Slack, calendars, or data warehouses. Kath explains that Codex infers intent and may suggest plugin connectors automatically, or you can explicitly request “bring your own connector.”
- •Codex tries to infer connector needs from the prompt and collaboration context
- •You can explicitly ask to use plugin connectors / “bring your own connector”
- •Sites’ value: handles deployment and “last mile” from local to web
- •Enables rapid internal tooling without bespoke API plumbing
- •Supports both prototypes and durable internal tools
- 7:53 – 9:00
Sharing Sites across teams: scoped access, sensitive data, and role-based views
Kath clarifies that sharing a Site doesn’t mean sharing the same underlying data. Each user’s authenticated connectors determine what they see, making Sites compelling for teams with sensitive or role-specific information like finance or data orgs.
- •Each viewer sees data based on their own auth + team context
- •Same Site can adapt to different teams (Sites vs Codex vs Identity, etc.)
- •Particularly useful for sensitive data and segmented responsibilities
- •Reduces risk of oversharing by keeping scope tied to the visitor
- •Creates a secure collaboration surface without key distribution
- 9:00 – 11:10
Favorite connectors and “personalized software” as a workflow pattern
Kath lists the connectors she repeatedly uses—Slack, Notion, Google Drive, and Google Calendar—and explains the spectrum of how people use Sites. Some build throwaway micro-apps; others replace rigid workflows with flexible, personalized tools that evolve with teams.
- •Most-used connectors: Slack, Notion, Google Drive, Google Calendar
- •~60 connectors available, with more being added over time
- •Use cases range from small prototypes to full internal tools
- •Sites enables personalized software that matches how individuals think/work
- •Kath builds temporary Sites for specific needs (e.g., business trip calendar views)
- 11:10 – 12:28
Sites as infrastructure: deploy, storage, data, and hosting plugins/evals
Kath reframes Sites as more than a place to host webpages—it can be infrastructure for building and running software. She mentions built-in capabilities like data/storage components, co-editing, and newly announced support for hosting MCP plugins via Sites.
- •Sites created to remove DevOps friction of deploying Codex-built software
- •Used for websites, plugins, hosting data, and running evals
- •Built-in functionality includes D1 data, R2 bucket, deployability, co-editors
- •New: host MCP plugins through Sites and reuse them in Sites workflows
- •Creates a “virtuous cycle” where Sites is both app surface and infra layer
- 12:28 – 16:20
Fixing a kid-ruined Spotify algorithm: a weekly playlist Site automation
Switching to consumer fun, Kath explains how her household’s music habits broke her Spotify recommendations. She built a mobile-friendly “heavy rotation” Site that scrapes popular playlists (via Reddit + computer use), then generates a fresh playlist every Monday morning.
- •Motivation: Spotify overwhelmed by kids’ content (Bluey, Sesame Street)
- •Uses instruction MD files to shape UI (mobile-friendly, simple)
- •Uses Reddit “top playlists” discovery via computer use
- •Automation runs every Monday at 8:00 AM to build a weekly playlist
- •Site embeds Spotify playback and can reference local/Apple Music libraries
- 16:20 – 20:25
Dungeon crawler in Sites: skills-based creation and community-built rooms
Kath demos a large user-created dungeon crawler and introduces the idea of distributing a ‘Dungeon Site’ skill so others can generate new rooms. The workflow encourages people to create themed rooms (e.g., Stranger Things) and share them for inclusion in an expanding dungeon.
- •Game built to showcase Sites sharing via URL and community remixing
- •Players choose characters (including Codex pet) and explore rooms
- •Dungeon Site skill provides constraints: room dimensions, controls, entities
- •Creators can instruct Astra to generate themed rooms (moon, volcano, etc.)
- •Vision: collect user rooms and attach them into a “largest” community dungeon
- 20:25 – 25:53
Analytics + growth: lightweight metrics today, organic sharing via hashtags
Claire asks how creators can tell whether people are using or extending their Sites. Kath notes analytics are currently basic (visits/traffic) and hints at ambitions for more robust, investigative, agentic analytics—while personally favoring organic community discovery via Twitter tags.
- •Current developer analytics are lightweight (primarily traffic/visits)
- •No direct insight into who is using a skill to build extensions
- •Future ambitions: more robust analytics, investigations, agentic analysis
- •Community strategy: hashtags/mentions on Twitter rather than centralized tracking
- •Preference for meeting builders and growing organically vs heavy instrumentation
- 25:53 – 30:06
Bringing inference into Sites: the widget experiment and “player-funded” creativity
Kath describes experiments to embed inference directly inside Sites so users can modify experiences in-place instead of bouncing back to ChatGPT. They imagine scenarios like players (or kids) having an allowance/budget to spend on in-game generation and edits.
- •Goal: enable ‘design mode’ inside the Site with embedded inference
- •Early experiment: a simple Site widget to prompt and debug from within Sites
- •Envisioned: game players bring their own inference to generate content live
- •Budgeting concept: allocate usage/spend (e.g., kids’ in-game generation allowance)
- •Ties to broader theme: real-time, interactive generative experiences on the web
- 30:06 – 31:56
Awesome Sites gallery: community showcase at awesomesites.ai
Kath explains how she created “Awesome Sites,” inspired by ‘awesome lists,’ to showcase what the community is building. Submissions come via #awesomesites and DMs, and the gallery highlights diverse projects like instruments, transit atlases, and games for inspiration.
- •Created a curated directory modeled after “Awesome Lists” culture
- •Submission workflow: #awesomesites, DMs, and community outreach
- •Showcases diverse examples (piano, transit/bus atlas, games, simulated worlds)
- •Positioned as inspiration for quick trip-specific or task-specific Sites
- •URL shared: awesomesites.ai
- 31:56 – 35:40
Prompting under pressure, AI boundaries, and wrap-up (where to find Kath)
In the closing segment, Kath shares her prompting style during demos—polite, but firm when the system makes unwanted assumptions. She draws a clear boundary against the model speaking in her voice (emails/Slacks), then shares her handle and invites feedback.
- •Prompting style: ‘please/thank you,’ but stern when behavior is wrong
- •Strong boundary: don’t let the model email/message people as her
- •Uses Codex heavily for research and drafting, but not final voice/authorship
- •Dot names and personal details, then contact info and feedback invitation
- •Kath’s handle: simpsoka; DMs open; Claire closes with subscribe/review asks
