She Ordered the Lobster. Now She's My Wife. — anatomy of a 1-hour AI video

The video above is 25 seconds long. It took me 62 minutes to make. There's no camera, no actor on a couch, no studio. The whole thing came out of Kinova Studio — the AI video agent I'm building — and I want to walk you through what actually happened, not the sanitized version.

I'm writing this because every time I see a “made with AI” demo, the workflow gets collapsed into a single screenshot of a final product, and I leave with no idea whether it survives contact with reality. So this post is the receipt. Times, credits, every regeneration, the dead ends, the model A/B I ran mid-session, the moment I argued with my own product about the safety filter.

If you're a faceless-channel creator wondering whether agent-based video is actually usable yet, this is for you.

What I started with

Two things: a script I wrote on my phone, and a character I'd built a few weeks earlier.

The character is Jay — a preset I keep around for first-person mindset content. Late 20s, Austin, sits on a light gray couch in a sunlit apartment, black henley, wooden bookshelf with a monstera behind him. Reads Marcus Aurelius. Drinks black coffee. Runs 3 miles before sunrise. The kind of guy who'd give you real talk over a beer and you'd think about it for a week.

I'm not pretending this is a small detail. The single biggest unlock for faceless video is having a defined character — one whose face, room, lighting, and clothing are written down somewhere a model can reach. Kinova stores this as a structured visual description, and every scene image generation pulls from it as the identity anchor. Frame-to-frame consistency stops being something you fight for; it just happens.

The script was a 4-scene rough draft — a quick visual note, the voiceover line, and an environment tag per scene. Restaurant setup, the check, walk to her car, twist reveal at home. 25 seconds total.

I pasted it in at 22:58 on a Wednesday.

The first pivot (and why it mattered)

The agent took my script and converted it to a structured scene plan in about 56 seconds — start-frame descriptions, camera moves, environment IDs, voiceover text per scene. Clean output. I read it.

Then I noticed it had built a scene plan with two characters in it: Jay and his date. Which is what my script literally said. But it's not how I actually shoot Jay.

So I told it:

for jay, i normally just let him sit there and tell the story, no other actors in the scene plan

Forty seconds later, it had reworked the plan. Jay alone on his couch, telling the entire story direct to camera. Same beats, same VO, same environment carrying through all four scenes. Just him.

This is the part that doesn't show up in product demos. The script-to-plan step isn't a one-shot generation; it's a conversation. The first plan was the literal interpretation. The second plan was the correct interpretation — the one that matched how Jay-content actually works on this channel. I wouldn't have known to instruct that at the start, because the issue only became visible once I read the structured plan back.

Lesson I keep relearning: the agent's first output is a draft for me to react to, not an answer.

Images first, then clips — and why I regenerated 3 times

Once I approved the scene plan, the pipeline moved to image generation. Four scene images — frame-1 of each clip, generated as stills first.

This is the most important architectural choice in the whole pipeline, and it took me a while to internalize: always generate the still image first, then animate it. Skipping straight to text-to-video is faster but you lose the ability to iterate on framing without burning clip credits. Images are cheap to regenerate. Clips are not.

Here's what the credit log looks like for this session's images:

• 1 initial generation (all 4 scenes): 15 credits
• 3 single-scene regenerations: 5 credits each → 15 credits

Three regens across four scenes, mostly to tighten Jay's framing — get the medium close-up sitting at the right distance, get his expression matching the scene's emotional beat. Dry-amused for the setup. Slow head-shake at the check. Genuine softness for the wife reveal at the end.

The images I picked were the ones that made me feel the scene before any motion was added. If a still didn't feel right, I knew the clip wouldn't either. Reroll the still. Don't pay for a clip that's wrong from frame one.

The Veo3 Fast vs Veo3 Pro question

This is probably the part most people want to know about, so I'll be specific.

Kinova lets me pick the underlying video model per clip. For this session I had two options: Veo3 Fast and Veo3 Pro. Same input image, same motion prompt, different model.

I A/B'd both on scenes 1 and 2 — the opener (Jay leaning forward into the lens, dry setup) and the check moment (subtle reaction, slow head turn).

What I observed:

Veo3 Pro rendered Jay's micro-expressions cleaner. The slight eyebrow raise, the half-smile that lands a beat after a punchline — these came through more clearly. Lip sync was tighter. The motion felt directed rather than drifted.

Veo3 Fast was a fraction of the cost and totally usable for scenes where the motion was subtle and the camera didn't move much. Talking head, light push-in, no body shift. Honestly, on a phone screen at 9:16, the difference is hard to see for the middle scenes.

My final cut ended up as 3× Veo3 Fast + 1× Veo3 Pro. The Pro went to the opening shot — scene 1, the first 8 seconds, the part that decides whether someone scrolls past or watches to the end. The hook gets the premium model. The rest can ride.

If you take one rule from this post: spend Pro credits on the hook, Fast credits on the body. That's the rule I'll keep.

Total clip credits for this session:

• Initial clip generation (all 4 scenes): 110 credits
• Regenerations across multiple scenes: 72 credits

The regens were mostly Pro re-rolls trying to land the opener, plus one re-roll of the wife-reveal at the end where I wanted the wedding ring to land in shot on the right beat.

The safety filter conversation

Around 23:30, mid-clip-regeneration, I noticed too many prompts were getting flagged. So I did something I think more people should do — I just asked the agent why.

why is my scene description easily triggers safety filter? could you search for sensitive words?

It told me, in its own product, the things it knew were likely tripping the underlying video model:

• “heels” or suggestive clothing references
• “wedding ring” + intimate framing
• “eye contact” + close-up framing
• (“market price lobster” — totally fine)

Then it did the useful thing — for the specific scene that kept getting flagged, it rewrote the prompt with cleaner wording. I re-triggered the regeneration. This one passed through clean.

Two things from this. First, the filter triggers are real, and they're triggered by descriptions the creator writes, not just by the visual content the model produces. If you write “she walks past him in heels with a wedding ring catching the light,” the underlying model will reject it regardless of how innocuous the actual frame would have been. Word choice in the prompt is downstream of word choice in your script.

Second — and this is the part that sold me on agent-shaped tools — the pivot I made for creative reasons earlier (Jay solo, no other actors) is also the lower-friction pattern for safety filters in general. Solo character, defined set, no romantic framing, no suggestive wardrobe — low-trigger by default. Good creative decisions are often also good production decisions.

The numbers

Here's the full receipt for the session:

What I'd do differently next time

Three things, written so I don't have to relearn them.

Why I'm building this

I started Kinova Studio because I'd watched too many smart creators bounce off raw video models. Veo3 and Sora are extraordinary, but they give you 8 seconds of clip, not a video. You still have to:

• Write a script
• Break it into scenes
• Keep a character consistent across those scenes
• Anchor environments and outfits
• Generate stills first to control framing
• Pick a model per clip
• Generate voiceover
• Sync VO to the clips
• Survive the safety filter
• Stitch it all together

That's a workflow, not a model. And right now almost nobody is shipping a tool that owns the whole workflow with a creator's instincts in mind. So I'm building one.

This video — Jay on his couch, dry as a martini, telling a 25-second story about a first date — is what 62 minutes of an agent and a creator looks like when the workflow actually fits. No camera. No studio. No second take. Just Jay, sitting there, telling the story.

Three years later. She's his wife.