Nick Baumann, a product lead at OpenAI, joined Lenny's Newsletter host Claire to demonstrate the ChatGPT stack he actually uses day-to-day — not polished demos, but real workflows running in production. The session covered Codex, ChatGPT Voice, ChatGPT Sites, and Heartbeat agents, and the cumulative picture is more operationally capable than most people following AI product releases probably realize.
Codex Voice: Parallel Tasks While You Keep Talking
The most immediately striking demo involved ChatGPT Codex Voice, triggered via a hotkey. Once active, the orb reads the current screen state through Appshots and can spin up separate threads simultaneously — in the live demo, Baumann had it booking a flight, filing an expense report, and checking his calendar at the same time, while he continued talking.
One underappreciated feature that Baumann highlighted: thread-forking. When a task branches into complexity, Codex can open a new thread, inherit context from an existing one, or pull from previous conversations on its own — no manual organization required from the user. For anyone who has wrestled with context windows or manually restructuring AI conversations, that's a meaningful quality-of-life shift.
Always-On Agents via Heartbeats
Heartbeats in ChatGPT Work are OpenAI's approach to background monitoring agents. Baumann set up notifications for:
- Suspicious charges appearing on linked accounts
- Package delivery proximity alerts
- Emails flagged as requiring attention
Critically, he built all of this without writing code. For startup founders and operators who want ambient AI monitoring but don't have engineering time to spare, this category of tooling is worth watching closely.
ChatGPT Sites: Closer to a Real Deployment Platform Than a Toy
ChatGPT Sites drew particular attention for how fully-featured it turned out to be. During the episode, Baumann and Claire built and published a live "How I AI" tips site — organized by tool and job function — in roughly the time it would normally take to configure a CMS template.
Under the hood, ChatGPT Sites includes:
- A SQL database
- File storage
- Environment variables
- Email-based access controls
For internal tools, lightweight apps, or gated resources, that feature set starts to look genuinely competitive. The comparison to low-code platforms like Webflow or headless CMS setups is becoming less flattering for those incumbents as this kind of AI-native deployment layer matures. Founders thinking about an AI startup website or internal tooling layer should note how quickly the cost and time assumptions around web deployment are shifting.
Video Editing Overnight, Hands-Off
Baumann's video workflow is perhaps the clearest illustration of AI eliminating real production labor. His process:
- Record 50–60 raw clips
- Upload to ChatGPT
- Talk through the rough story structure
- A custom plugin transcribes everything, scans footage, selects the strongest takes, and assembles finished vertical videos — overnight
What would typically occupy most of a content team's day is largely done while he sleeps. For solo operators or lean marketing teams producing short-form video, that's a workflow worth reverse-engineering.
Voice as a Context-Delivery Mechanism
A recurring theme across the session: voice input produces better AI outputs than typed prompts, not because the model processes audio differently, but because people give richer context when they talk freely. Claire dubbed this the "yapper's API" — the idea that removing the pressure of crafting a perfect prompt lets users convey more nuance, which in turn produces more useful results.
This pattern appeared across voice chat, mobile dictation, and the Codex orb interface consistently.
Latency as a Product Decision
Baumann flagged latency as an increasingly strategic variable. Once a model has the right tools, raw reasoning quality has diminishing returns — speed becomes the differentiator. Many AI products are still designed around the assumption that waiting is unavoidable. As that changes, tools that feel immediate will have a compounding advantage, especially for users who currently find AI too slow for live, in-the-moment work.
The full workflow walkthroughs are available at chatprd.ai/how-i-ai/nick-baumanns-3-advanced-chatgpt-workflows.
