For the first six of these experiments I used AI to make the videos. This time I turned it around: I had an AI interview me. A real voice, in a real meeting, asking real questions — while I stood out in the garden.
There’s a thirty-second version too, for the punchline: https://www.youtube.com/watch?v=JxX3ozLCXEk
What actually happened
I joined a Google Meet on my phone, walked out into the garden, and let an AI voice run the conversation. It opened, it asked, it followed up on the interesting bits, it summarised back what I’d said and pushed for more. For about nine minutes it behaved like a competent interviewer who’d done the reading. The experiment was simple: not an AI that transcribes a meeting afterwards, but one that participates in it.
This connects to something we’ve been circling at Rover Engineering for a while — the idea of a Robot Project Manager. Not a tool that replaces a project manager, but one that takes the coordination load: task tracking, risk, knowledge management, and surfacing the insight that actually needs a decision. The point isn’t to remove people. It’s the opposite. As I said in the interview: get the data flowing on its own, so an engineer spends their time on engineering instead of moving information around, and the genuinely nuanced, critical-thinking parts of running a project get the human attention they deserve.
The accidental thesis statement
And then, near the end, with the AI mid-way through a thoughtful summary of everything I’d said, I noticed one of our sunflowers had snapped. I apologised, told the AI I had a garden emergency, thanked it, and left. It said “no worries — good luck with the garden,” and that was the end of the meeting.
I’ve left that in, because it’s the whole idea in one accidental moment. The meeting ran itself, so the human could go and deal with the thing that actually mattered right then. That’s the future I find genuinely interesting — not AI doing the human stuff, but AI holding the structure so the humans are free for the parts that need them.
How this one got made — and what broke
This is the seventh in an open series documenting how these videos get produced, and testing how far AI and automation can take real video work. The new capability this round was the big one: an AI as an active, spoken participant in a recorded meeting, with the recording then run through the normal pipeline — analysis, transcript, edit, framing narration, b-roll, captions.
In the spirit of the series, the honest notes matter as much as the result:
- The interviewer was ChatGPT’s voice (OpenAI), brought into the call. I’d actually built a dedicated ElevenLabs interviewer agent with a Claude brain for exactly this — named it, gave it a system prompt, wired up the audio routing — but on the day it was simply faster to bring ChatGPT’s voice into the Meet. The purpose-built agent is sitting ready for next time, which is its own small lesson: the elaborate setup isn’t always the one that ships.
- The analysis model got the timestamps wrong. Gemini’s first-pass transcript drifted past the actual end of the recording — it placed the ending well beyond where the video stopped. If I’d trusted it, every cut would have been off. So the edit was timed from a Whisper transcript instead, and the cut points snapped to real phrase boundaries so nobody gets clipped mid-sentence. A good reminder that “the AI gave me timestamps” is the start of the job, not the end.
- The Robot Project Manager pitch footage appears throughout as muted b-roll — the abstract shots over the parts where I’m describing the vision — while the conversation audio carries on underneath.
The full pipeline this round: ChatGPT (the AI interviewer voice), Google Gemini 2.5 Pro (video analysis), Whisper / faster-whisper (transcript and captions), ElevenLabs (my cloned voice, for the short framing narration top and tail), Pillow and ffmpeg (framing the phone footage, lower-thirds, the b-roll, ducking the music under the dialogue), and Claude Code orchestrating the whole thing. The marginal cost was pennies.
Where the series is
If you’re following along: experiment one set the baseline, then it grew — captions and a Final Cut export, AI b-roll, user footage and music, my own voice in the cut, and a live product launch. This is experiment seven: the AI stops being the subject of the video and becomes a participant in making it.
The pattern that keeps showing up is the same one the sunflower made literal. The machinery is increasingly happy to run on its own. The interesting question is always what the people do with the time that frees up.
Rover Planet — fuel your curiosity.
