How I AIHow this OpenAI engineer uses Codex + ChatGPT Work to automate everything
CHAPTERS
- 0:00 – 3:47
From chatbots to coding agents: why Codex feels like a step-change
Nick frames the shift from early ChatGPT as a conversational chatbot to modern coding agents that can read/write files, run commands, and autonomously figure out intermediate steps. This sets up the episode’s theme: delegation and automation as the new interface.
- •Early ChatGPT made AI feel accessible as a conversational partner
- •Coding agents expand capability to real actions (files, commands, workflows)
- •Autonomy: you give one instruction; the agent decomposes and executes subtasks
- 3:47 – 5:52
What’s new in Codex: threads, forking, and an orchestration layer
Nick and Claire discuss under-discovered Codex/ChatGPT primitives—creating new threads, messaging existing ones, and letting threads coordinate. Voice becomes the “orchestration layer” that makes these capabilities easier to use and discover.
- •ChatGPT can create/fork threads and message existing threads
- •Threads can coordinate, helping manage context across tasks
- •Voice + screen context turns these hidden primitives into a fluid workflow
- 5:52 – 7:01
ChatGPT Work + Heartbeats: always-on monitoring via connectors
Nick explains why he uses ChatGPT Work (especially on mobile): connectors/plugins plus compact context make it powerful for day-to-day automation. Heartbeats proactively notify him about email, purchases, packages, and anomalies without constant checking.
- •ChatGPT Work mirrors the “agent harness” experience
- •Connectors (e.g., Gmail/GCal/finance tools) enable real monitoring
- •Heartbeats deliver proactive notifications a few times per day
- •Automation reduces attention load—no need to constantly watch inbox/finances
- 7:01 – 7:35
Live Voice demo: hotkey, screen awareness, and task delegation
Nick triggers Voice via hotkey and demonstrates it reading on-screen Slack context (the Paris offsite details). He then delegates booking flights and hotel planning, showing how Voice can spin up background tasks and coordinate them.
- •Hotkey launches a floating Voice interface anywhere on desktop
- •Voice can see the screen and interpret open apps (Slack)
- •Delegates travel planning: check calendar availability + prepare options
- •Spins up separate tasks/threads that run in the background
- 7:35 – 11:13
Under the hood: AppShots, browser sessions, and multi-thread control
They unpack what just happened: AppShots act like “smart screenshots” loaded with context; Voice effectively has this built in. Nick shows directing tasks to use his existing Chrome session and to stage actions without submitting until confirmation.
- •AppShots: context-rich captures of the active window
- •Voice can manage tasks and send updates to running threads
- •Control over execution: use existing Chrome session, prepare but don’t book
- •Practical delegation: parallel tasks for travel + expenses
- 11:13 – 15:23
Voice UX debate: latency vs. intelligence (and why planning works better aloud)
Claire and Nick debate whether fast responses or smarter reasoning matters more for Voice. Nick prefers high intelligence for voice, but notes that with strong tools/delegation, latency becomes the key differentiator; both predict product design will shift as latency shrinks.
- •Voice enables better “thinking flow” than dictating long text blocks
- •Tradeoff: high intelligence vs. low latency
- •If tools can be delegated to reliably, latency matters more
- •As latency drops, AI product design assumptions will change
- 15:23 – 16:58
Mobile voice workflows: maintaining a living site from your phone
Nick describes a real mobile workflow: a ChatGPT-managed site that aggregates Codex/ChatGPT usage examples. He drops links or dictated prompts into a thread, and ChatGPT Work updates, edits, and deploys the site—entirely from mobile.
- •ChatGPT thread manages a running website project end-to-end
- •Add items by simply pasting links and saying “add it”
- •Prompts tab used to collect/share reusable workflows
- •Editing and deployment can happen from ChatGPT Work on mobile
- 16:58 – 18:35
ChatGPT Sites explained: database, storage, privacy controls—and building live
Nick explains ChatGPT Sites in both technical and non-technical terms (SQL DB, S3-like file storage, env vars). They then kick off a live build: a ‘How I AI’ tips site that pulls best moments from Claire’s videos and enables filtering by tool and function.
- •Sites include a SQL database, file storage, and environment variables
- •Non-technical framing: a controllable website that can store and organize content
- •Privacy options: public or restricted by email
- •Live build brief: branding + filters by AI tool and audience/function
- 18:35 – 21:36
Two-person voice capture: the ‘yappers API’ for requirements gathering
While the site runs, Claire shares a tactic: using speakerphone to capture someone else’s requirements into ChatGPT via Voice, turning unstructured rambling into actionable output. Nick echoes this with an internal workflow: get messy voice/context instead of scheduling meetings.
- •Voice enables “two-person” requirement capture via speakerphone
- •High-bandwidth context beats structured docs for many tasks
- •Replace meetings with voice notes + agent interpretation
- •Unstructured input is acceptable if the agent can decipher it
- 21:36 – 26:46
Live UGC editing in Codex/ChatGPT Work: dump clips, extract story, assemble cuts
They pivot to creator workflows: Nick drags many short clips into ChatGPT Work and uses a custom ‘UGC Video’ plugin to generate a vertical UGC-style edit. The process involves transcription, visual understanding, best-take selection, and assembling a narrative automatically.
- •Workflow: upload many clips → auto-transcribe → analyze visuals → pick best takes
- •Voice/dictated narrative helps the model understand the intended story arc
- •Handles common creator friction: you don’t need to pre-select good takes
- •Outputs multiple aspect ratios (e.g., 9:16 and 4:5)
- 26:46 – 28:30
Plugin creation for repeatability: steering, safety zones, captions, and delivery formats
Nick explains that the ‘UGC Video’ plugin encodes hard-won steering from previous runs: IG safe zones, caption conventions, breakpoints, and format requirements. He built it by turning a successful long thread into a reusable plugin via an in-app plugin creator skill.
- •Plugins capture bespoke process knowledge for consistent outputs
- •Encodes constraints: ad safe zones, text placement, caption style, breakpoints
- •Built from a ‘massive thread’ plus tweaks for a reliable happy path
- •Reduces back-and-forth questions and standardizes deliverables
- 28:30 – 31:45
Redaction and blurring: automatically protecting sensitive info in screen recordings
Nick shows output that includes automatic blurring of addresses and even fine-grained redaction that tracks while scrolling. They discuss how manual redaction is a non-starter for creators and teams; letting the model iteratively verify blur quality removes a major publishing barrier.
- •Automatic blur/redaction can track sensitive elements across frames
- •Model verifies its own redactions through iterative checks
- •Solves a major pain point for YouTubers and internal demos
- •Also reduces the mental overhead of planning “safe” recordings up front
- 31:45 – 32:29
Speed and parallelism with SOL Ultra—and shipping faster with agent workflows
Claire asks whether SOL Ultra is needed; Nick says yes primarily for speed and parallel micro-tasks (frame checks, transcript processing, judgments). He describes recording multiple videos, uploading dozens of clips, dictating narratives, and returning later to ready-to-ship edits.
- •SOL Ultra: multi-agent framework enabling parallel processing and judgment
- •Faster handling of many micro-tasks in video workflows
- •Batch workflow: record a lot → upload 50–60 clips → dictate narratives → review later
- •Lowers barrier to consistent content production
- 32:29 – 35:02
Results: the How I AI site preview + sharing and access controls
They review the generated How I AI site and notice how it organizes episodes by use case (with some duplicates to refine). Nick explains publishing and access: share publicly or restrict by email, with viewers logging in via ChatGPT.
- •Site matches requested branding (purple/black/white)
- •Content organization by use case; room for iteration (duplicate entries)
- •Can share publicly or whitelist specific emails
- •Authentication is “Log in with ChatGPT”
- 35:02 – 41:34
Lightning round: voice in offices, ‘feeling AGI,’ and prompting when it goes wrong
Nick and Claire close with rapid questions: voice is best at home but awkward around coworkers, pushing people into pods/rooms. Nick’s ‘feel AGI’ recommendation is ChatGPT Work + connectors + Heartbeats as an accessible always-on agent; they also joke about escalating prompting styles as expectations rise.
- •Office reality: voice is awkward; people adapt with pods/meeting rooms
- •‘Feel AGI’ moment for mainstream users: Work tab + plugins + proactive Heartbeats
- •Agent harness makes autonomy accessible beyond developers
- •Prompting under failure: redirecting (sometimes with frustration) as expectations rise
