How I AI: Nick Baumann's 3 Advanced ChatGPT Workflows for Voice Automation, Site Building, and Video Editing

If you've ever wished for a personal assistant, the workflows Nick Baumann from OpenAI demonstrated in this episode of How I AI are a glimpse of that future. He showed how to delegate complex, multi-step tasks like coordinating travel and filing expenses using a single voice command, building a live website from a spoken description, and turning a folder of raw footage into an edited vertical video.

We tested three of his advanced workflows: managing personal logistics with ChatGPT Voice, building a searchable How I AI site with ChatGPT Sites, and editing video with a custom UGC Video plugin. See Automate UGC Video Editing From Raw Clips with ChatGPT.

Coordinate travel and expenses with a single voice command

Nick uses ChatGPT Voice as a personal assistant to coordinate travel, check his schedule, and prepare expenses, all while keeping multiple tasks running in parallel. You can follow the full implementation in Use ChatGPT Voice as a Personal Assistant for Logistics and Travel. See Use ChatGPT Voice as a Personal Assistant for Logistics and Travel.

Give the AI context by sharing your screen

Nick triggered ChatGPT Voice with a hotkey that works from anywhere on his computer. The key is that the AI can see what's on screen. For the demo, he had a Slack thread open about a team offsite in Paris. Instead of copying links or summarizing the details, he just started talking, letting the system capture what he calls an "app shot", a context-rich screenshot, to understand the request.

The system captures what Nick calls an app shot, a context-rich screenshot the model can inspect before acting.

Delegate the travel search and calendar check

With the offsite context loaded, Nick switched into delegation mode. He spoke naturally, asking the AI to book flights and a hotel, but also to check his calendar to see if he could extend the trip through the weekend.

I'm hoping you can help me book my flights and my hotel for the Paris offsite. I'm also thinking about sticking around through the weekend and maybe a couple days after. Could you maybe like check my calendar, see what my availability's like, and see if it's, if it's feasible for me to stick around?

That single voice command kicked off a chain of actions. The AI used plugins to check his calendar for conflicts, then looked at travel options through Navan. It identified a meeting that would conflict with a Tuesday return and suggested coming back on Monday instead, working through the tradeoffs of extending the trip just like a human assistant would.

Start the expense task in parallel

While the travel agent was running in the background, Nick demonstrated the system's ability to manage parallel tasks by starting a completely separate thread for an expense report.

While I've got you, I've got an email, I think. There should be some Amazon receipt for a mic I just bought recently. It's for DX uses. I'm hoping you can help me expense it. Could you spin up another thread to handle taking care of that expense report for me?

The new task searched his email for an Amazon receipt for a microphone and staged an expense report in Navan. Crucially, it waited for his final approval before submitting. Voice acted as the control surface for two independent, active tasks.

Build a searchable episode archive with ChatGPT Sites

Next, Nick demonstrated ChatGPT Sites, a feature that can build and deploy a live website on a chatgpt.site domain from a natural language prompt. He explained that the platform is surprisingly robust, supporting a SQL database, S3 file storage, environment variables, and access controls. See Build and Deploy a Live Website with a Single ChatGPT Prompt.

To test it, we decided to build a searchable archive of tips and workflows from How I AI.

Describe the site

I described the requirements out loud, and Nick translated them into a detailed prompt. The goal was a site that would pull the best tips from the show's YouTube channel, make them filterable, and provide links to specific timestamps. We also specified the How I AI branding, "good old AI blurple," purple, black, and white, and asked for categorization by tool and function.

I'm here right now with Claire Vo. She hosts the How I AI podcast, and we're hoping that we can build a site that takes all the best tips from the videos on her YouTube channel and makes it so you can like easily filter through them and have quick links to like different timestamps on her videos. Make it match the How I AI branding. We use good old AI blurple, purple, and black and white. And let's have it, if possible, categorized by tool... and also function... Otherwise, I'm excited to see what you come up with.

Generate the site and its data structure

With the prompt submitted, the system went to work. It designed the necessary data structure, built a user interface, and began populating the site by pulling material from the How I AI YouTube channel and organizing it into the requested categories.

Review the access settings and publish

The resulting site used the requested purple, black, and white color palette and presented a filterable gallery of episode workflows. It correctly identified different use cases within single episodes and created separate entries for each, making the archive genuinely useful.

Publishing is flexible. You can make a site public or restrict access to specific email addresses. Authentication is handled through a user's ChatGPT login, which makes this feature particularly useful for building internal tools, lightweight dashboards, or private MVPs that you want to share with a small group.

Automate video editing with a custom UGC plugin

For his final demo, Nick showed a workflow for editing user-generated content (UGC) style videos. The process automates take selection, assembles a vertical video, formats captions, and even finds and redacts sensitive on-screen information.

Add the raw footage

Nick started by dragging a folder of raw video clips directly into the ChatGPT interface. The footage was from a recent park shoot and contained multiple takes, rough cuts, and filler, the kind of material that makes a first-pass edit so tedious.

Describe the edit through a custom plugin

He then dictated a prompt explaining the story he wanted to tell, using a custom UGC Video plugin he built. This plugin contains his preferences for video formats (9x16, 4x5), caption styles, and other stylistic elements.

I recorded a video for the record and replay feature. It starts with me just explaining that I'm apartment hunting... and then I walk through and show the feature in real life... Can you go through these clips first, like pull the transcripts, find the best takes, and then kind of piece this together into a UGC-style video?

Let the agent process and assemble the footage

Nick ran the edit using Soul Ultra, which he described as a framework for coordinating multiple agents and checking their work. This allows for parallel processing of the complex tasks involved in the edit:

1. Transcribe all clips to understand the available dialogue and story beats.
2. Select the best takes by analyzing the transcripts and listening for spoken cues from the creator, such as "cut," "bad take," or "good take."
3. Inspect key frames to understand what is happening visually on screen.
4. Assemble the selected segments into the requested narrative arc.
5. Identify and redact sensitive information with tracked blurs. Nick showed the system blurring street addresses and even specific lines of text in a Slack window, then checking and revising its own redactions to ensure they were accurate.

This workflow eliminates the most tedious parts of editing: reviewing every take and manually redacting sensitive content. The human editor still reviews the story and the final output, but the system handles the time-consuming first pass. As Nick noted, this dramatically lowers the barrier to entry for creating short-form content.

What's worth copying

Nick's demos all point to a more powerful form of delegation. Instead of breaking down a task into small, sequential steps, you can provide high-level intent and let the system manage the execution. Voice provides the initial context, parallel agents keep administrative work moving, ChatGPT Sites turns a spoken spec into a functional tool, and the video plugin handles a complex creative task in the background.

These workflows are worth copying for anyone who spends too much time on administrative setup or tedious first drafts. The voice-driven multitasking is the most immediately useful; being able to kick off an expense report while a travel search runs is a real time-saver. The site builder is perfect for quick internal tools or MVPs, but a production site will still require human design and development. Similarly, the video editor is a powerful rough-cut machine, but the final storytelling and polish still depend on a human's creative judgment. The real shift is in how work gets started: one high-context conversation can now launch and coordinate complex work across multiple tools, even if a human still needs to approve the final result.