On 5 August I wasn’t building anything yet. I was just trying things out.
I stood one of my kids’ plasticine figures on the kitchen table and took a photo with my phone. Then I wrote a single sentence: come to life from this state and greet the children by name. Nothing else. No framing, no lighting, no style, not a word of English.
It came out right first time. Ten seconds of the figure moving on my table, with my things in the background, saying the names.
Five days later I was running twenty-five takes in a row without a single good one, using two-hundred-word requests in English with their list of prohibitions. That material is the next report. Here I’m writing down what I already had in front of me and failed to read.
The photo was doing the work
The sentence described nothing, and it didn’t need to: the table, the kitchen light, the size of the figure and the fingerprints in the plasticine were already in the image.
That’s data I don’t know how to write down. The phone camera gives it away for free.
When I later started explaining in writing what could already be seen in the photo, the result got worse. It’s counterintuitive and it took me a week to accept. But describing the image you’ve attached confuses the model instead of helping it, because you’re handing it two versions of the same thing and one of the two is worse.
There was a second thing, and that one was my decision. Before asking for any clip I wrote that I preferred the plasticine characters to come to life in their own surroundings.
That sentence fixed the set for everything that followed. So the look isn’t decided shot by shot: it’s decided once and inherited.
Retrying beats rewriting
When a clip came out so-so, my instinct was to rewrite the request. I tried the opposite. I asked for a better video without touching a comma, and it came out better.
It’s the same model with the same instruction, so the only thing that changes between one take and the next is the roll of the dice. Now I repeat first. I only touch the text if repeating fails twice.
What can’t be done
I tried three times to continue from the previous clip, uploading it so the scene would carry on from there. All three died with the same message: editing uploaded videos isn’t available in my region.
So continuity between shots isn’t done with the previous clip, but with a photo. And that’s where the problem in the next report starts.