CHAPTERS
Goal and setup: cloning Claire with Google Flow + Gemini Omni
Claire sets the challenge: create an AI video avatar of herself and turn it into a ~1-minute hype video for the How I AI podcast in about 15 minutes. She opens Google Flow and explains the promise of Gemini Omni for consistent-character video generation, noting it didn’t work well on launch day but might now.
Avatar capture on phone: QR scan, photos, head turns
She starts the avatar creation flow by scanning a QR code and using her phone camera to capture required angles. The process includes taking multiple photos and turning her head left/right to complete the scan.
Prompting Flow as a “creative suite”: brainstorming the hype video concept
With an avatar available, Claire asks Flow to help storyboard a hype video for How I AI. She provides creative direction (dark green home office, books/posters, authentic but high-tech “hacker vibe”) and frames this as AI unlocking new creative capability for non–video creators.
Storyboard generated: seven-scene structure and visual beats
Flow proposes a seven-scene storyboard: keyboard close-up, office wide shot, chair spin reveal, HUD/montage moments, lifestyle beat, mic/podcast moment, and title/CTA. Claire reacts to the cheesiness but approves the structure and moves to generate scenes with her avatar reference.
First scene generation attempt: environment details leak into the avatar
Claire tries to generate the first scene using the storyboard prompt while @-mentioning her avatar. She notices the model pulls in background elements from her capture photos (posters/books), which is surprising and helps continuity—though it also reveals how tightly the avatar ties to the training context.
Troubleshooting mistake: generated images instead of videos
She realizes she accidentally ran image generation rather than video generation due to a UI toggle. After correcting the setting, she re-submits the prompt to generate actual video outputs, noting that video takes longer and typically produces two variants (V1/V2).
Batching seven scenes: waiting, variants, and jump-scare realism
Claire queues all seven scenes for generation, describing the experience as both exciting and uncanny. She reviews early outputs—like unexpected blue nail polish and a startling chair-spin clip—highlighting how close the avatar can feel while still producing surprising details.
Scene review: the chair-spin intro line and character consistency quirks
She plays two versions of the chair-spin scene where the avatar says, “I’m Claire, and this is How I AI.” The background includes a surprising NVIDIA reference; one version spins twice, and she chooses the one that flatters her more while noting hair and styling differences across clips.
Stitching in Flow: browser timeline editor to assemble the hype video
Claire demonstrates Flow’s built-in web editor timeline and stitches together her preferred takes in the suggested order. The assembly takes about five minutes, emphasizing an end-to-end pipeline from capture → storyboard → generation → edit within a single tool.
Premiere: the finished “How I AI” hype video
She plays the completed one-minute hype video, featuring the hook “We were told AI would replace us,” an intro, a montage of AI-themed visuals, and a subscription call-to-action. The result feels surprisingly polished for the time invested, despite occasional glitches and awkward phrasing.
Postmortem: what worked, what didn’t, and why it’s still impressive
Claire breaks down strengths—speed, low effort, and a ~50% “there” result that’s shareable—and weaknesses like uncanny facial moments, inconsistent backgrounds, and early-2000s “AI aesthetics.” She notes the avatar can look highly accurate in some frames and less so when expressing emotion (e.g., laughing).
Closing thoughts: experimentation, future improvements, and viewer prompt
She concludes that with tighter prompting, more consistent background constraints, and additional reference images, the results could become convincingly realistic. Claire invites viewers to try creating their own avatars and share examples, framing the episode as a successful (and strange) How I AI experiment.
