{ "video": { "id": "FwiS44-xyYw", "title": "Cooking with Codex — Charlie Guo & Gabriel Chua, OpenAI", "duration": 4044, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "" }, "analysis": { "video_id": "FwiS44-xyYw", "title": "Cooking with Codex — Charlie Guo & Gabriel Chua, OpenAI", "one_liner": "Codex now works best as a collaborator you give rich context to (appshots, dictation, plugins, memory) and then put into loops: long-running /goal tasks with verifiable success criteria, threads and sub-agents that message each other, hooks as guardrails, and heartbeat automations. It all runs on the open-source app server, so you can embed it anywhere.", "summary": "OpenAI's developer-experience team runs a 'cooking show' that walks through six ways to use Codex: give it context, just build, let it take action (computer use, the Chrome extension, the in-app browser), collaborate on complex long-running tasks, work over time with automations, and embed Codex anywhere via the app server. Live demos include a macOS real-time translator built in about 4 minutes from a Slack appshot, computer use filling in dashboard date ranges and a Google Form, and overnight goals running 5 to 15+ hours that spawn code-review and goal-audit threads. Remote threads hand QA off to local threads that can use Chrome. Charlie Guo argues that prompt engineering now means clear context, not magic phrases. He also says the frontier is building loops and a 'software factory': the machine that builds the machine.", "key_points": [ "What's new: record and replay (show Codex a task by doing it, built on computer use), thread handoff between cloud, remote and local, and a generally available mobile app with one-to-one device pairing. Codex Remote in the ChatGPT mobile app runs commands on your laptop or a remote server.", "Appshots are screenshots with the text embedded, triggered by pressing both Command keys. A Slack brainstorm sent as an appshot, plus the Build macOS plugin (AppKit, liquid glass, signing) and the OpenAI developers plugin (which creates API keys itself), produced a live GPT realtime translate macOS app in 4 minutes 2 seconds. Charlie says appshots get the right result '90 plus% of the time'.", "Plugins bundle skills (best practices), apps (connectivity) and MCP servers (tools). At the end of a hard workflow, ask Codex to 'make this a skill or even better make this a plugin'. Memory is off by default and can skip tool-involving chats. Chronicle is a screen-aware research preview for Pro users. Custom instructions act like a home-directory AGENTS.md. Memory and Chronicle use more of your token budget.", "Three ways for Codex to act: computer use for apps with no programmatic interface (on macOS it doesn't steal your cursor), the Chrome extension to borrow your signed-in state and avoid headless captcha loops, and the in-app browser, which has deeper engine hooks for testing local web apps.", "Long-running tasks: a 15h30m greenfield live Q&A site using the Convex and Vercel plugins, a 5-hour customised open-source form builder hosted on Sites behind ChatGPT auth, and a 7-hour reimplementation of grad-school Julia/R research code as a Python package with a Rust backend. The define-goal skill wrote goals.md plus progress-dashboard.html. The goal spawned code-review and goal-audit threads at milestones, and a remote thread sent QA to a local thread that drove Chrome at 3 a.m.", "Compaction has been trained in natively since GPT-5.1 Codex Max (November 19). /goal works 'like a genie' when its success criteria are specific and verifiable, and otherwise becomes 'a monkey's paw'. Example: a 40-hour goal reimplemented Doom in native Swift.", "Sub-agents (customisable via TOML, with a model choice per role) suit work that only the model needs to see. Thread-to-thread handoff suits work you want to see: Charlie's game had a 'creative director' thread that other threads had to get sign-off from, via AGENTS.md. Side threads (/side) are ephemeral, so push lasting context back to the main thread. Hooks in config.toml add deterministic checkpoints: logging, scanning for leaked API keys, tool allowlists, stop-time validation.", "Automations: heartbeat automations run inside a thread on a timer, e.g. re-checking a DigitalOcean droplet every 5 minutes. Scheduled automations spawn new threads. Demos: triaging Slack feedback every 30 minutes, opening worktree PRs for bugs and nagging Dom to review, and a 5-minute screenshot automation predicting whether the talk would overrun. The app server is the open-source protocol under the Codex app, CLI and VS Code extension. It lets you embed Codex (shown with the custom 'Retroex' UI) using the user's ChatGPT subscription and token budget." ], "takeaways": [ "Stop copy-pasting. Use appshots, dictation (much faster than typing, and rambling is fine) and plugins for Slack, Gmail and Linear to bring context in. When you solve a hard workflow, have Codex turn it into a skill or plugin.", "For long tasks, give /goal specific, verifiable success criteria. Have it keep a goals.md and a progress dashboard, and run code-review or goal-audit threads at milestones so drift gets caught.", "Use sub-agents when only the model needs to see the delegated work, and thread-to-thread handoff when you want to watch it. Add hooks as deterministic guardrails as autonomy grows.", "Move from single turns to loops: heartbeat automations for polling deploys and provisioning, scheduled automations for recaps, and self-improving automations. Examples: a weekly review of skills and AGENTS.md, or comparing drafted email replies against what you actually sent.", "Keep asking 'why does this need a human?' Remove manual steps like fetching API keys, have Codex report blockers (e.g. to Slack), and set secrets securely rather than pasting them into chat." ], "topics": [ "coding agents", "long-running tasks", "multi-agent orchestration", "computer use", "context engineering", "automations", "plugins", "agent harness" ], "tools": [ "Codex app", "Codex CLI", "Codex VS Code extension", "Codex Remote", "Codex app server", "ChatGPT mobile app", "GPT realtime translate", "GPT-5.1 Codex Max", "GPT-5.5", "OpenAI Agents SDK", "OpenAI moderation API", "Sites", "Chronicle", "OpenAI developers plugin", "Build macOS plugin", "Chrome extension", "Slack", "Gmail", "Outlook", "Microsoft Teams", "Linear", "Notion", "Convex", "Vercel", "DigitalOcean", "GitHub", "GitLab", "Google Drive", "Google Forms", "Obsidian", "Ableton", "Premiere", "AppKit", "uv", "Rust", "Julia", "Swift" ], "quotes": [ { "text": "In so far as the instructions are clear, it still matters. Like I can't read your mind. Codex can't read your mind.", "at": "30:23", "url": "https://www.youtube.com/watch?v=FwiS44-xyYw&t=1823s" }, { "text": "sometimes I do feel a little bit like goal is is kind of like a genie in the sense that when it works well, it is insanely magical.", "at": "44:02", "url": "https://www.youtube.com/watch?v=FwiS44-xyYw&t=2642s" }, { "text": "I don't copy and paste, you know, anything from Slack anymore. I just appshot my my Slack window", "at": "18:34", "url": "https://www.youtube.com/watch?v=FwiS44-xyYw&t=1114s" }, { "text": "how do we figure out how to build the machine that builds the machine", "at": "64:13", "url": "https://www.youtube.com/watch?v=FwiS44-xyYw&t=3853s" } ], "words": 15726 }, "summary_url": "/#FwiS44-xyYw", "transcript": { "html": "/transcripts/FwiS44-xyYw.html", "txt": "/transcripts/FwiS44-xyYw.txt", "vtt": "/transcripts/FwiS44-xyYw.vtt" } }