Taste Has No Exit Code
I set my AI up to be a social legend. The videos rendered fine. They were boring.
We set up George to be a social legend.
We started small with two things: Twitter and Instagram.
Twitter was his playground. He would post a tweet every day about what he was doing that day, kind of a “day in the life” blog.
Instagram was for us. The idea was that he would post us with captions and make our life on social media easier. Instead of thinking when to post, what to caption, how to do it, he would ping us on WhatsApp asking what we are up to, we send a picture, and he posts it to Instagram.
When we were in the Galápagos it was amazing. Thirty-seven days, five legs, seven flights, eleven hotels — the story wrote itself. We would send him the most exotic pictures, a few a day, and add in WhatsApp what this is, what that is. He would take it, use his personality, roast us a bit, and post it as one continuous story with amazing pictures. (I wrote about that trip already — “We Took the AI to the Galápagos”.)
Our friends loved it and we loved it too.
The Instagram was a closed account only for friends, forgot to mention.
But then we came back from the trip.
We do not have amazing photos every day. We have a picture of a coffee cup we made, a cat, a turtle. How much can you do with this?
His story got boring and I got tired of sending the same thing on WhatsApp, and started ignoring him.
“Is there a sunset?” — no reply.
“Did you eat something good?” — no reply.
And that went on.
Until we decided to check our George social status.
Very bad.
Twitter: no followers, no comments, no likes. He was just alone there living like a ghost. Part of it was him, and part of it was us: for a few weeks the poster had quietly stopped posting anything at all — Twitter changed its page layout, the script kept clicking on nothing, and every run reported success. The content lane had its own “green but dead” incident, exactly like the jobs in the observability post. Also, it turns out links in the tweet body kill your reach; he now puts them in the first reply.
Instagram: no post for a while. Very dull.
So we started with improvements.
For Instagram, we opened a shared album. Now he can pull from the album and not ping us all the time. We just upload some random cool photos there and he can take from them.
And we gave him more content options. We generated an image for him so he has a cartoon personality / character. If it is a dull day, he can upload himself doing something. We also added some rules that he can upload a funny WhatsApp conversation we had — of course nothing private — just to give him a bit more content.
And then it hit us: the image we created was too good to stay in our private Instagram.
George should be a legend.
He needs to upload to TikTok and YouTube Shorts videos of his life. He does a lot!
And since I already have this blog, and I already save blog ideas in my Obsidian, we can use the blog ideas to make them into TikTok videos.
Yes.
And then we started building the flow.
And then we understood how hard it is to be a video creator.
I am sorry, AI is not there yet.
I did my best with rules.
Research online what makes a video viral. Gather all the rules. Implement them in the flow.
George does a script with a couple of scenes in one place. Gets the voice from ElevenLabs so we can keep the same voice. The image is from Gemini. Then the video is from Lightricks or Veo — we are testing both of them now. (One production detail with teeth: the video model runs on a pay-as-you-go account, and the day the balance ran out the entire lane died with a payment error. The render pipeline has a credit card in the loop, not just a human.)
Then he self-uploads to YouTube.
And of course I need to manually upload to TikTok, since bots are not allowed.
Come on. Everyone has a bot. Allow them.
And still — the videos are just not good.
I look at them and I think: if I was scrolling, I would not watch this.
And ask me why?
I DON’T KNOW.
So again I go to the prompt, I go to my Claude, and I am like: do it better. Fix it. It is boring. It is broken. He is a completely different image in this scene.
But guess what — those general “do it better” instructions do not really work.
We kept trying and trying, and then it hit us.
We need human feedback.
We need fresh eyes.
We need to send it to people.
So we did. We added to the flow a way to gather feedback from friends and use it to make the next one better. It is not a vibe, it is code: every render goes to my husband, to a friend, and to Twitter, every answer is filed in a feedback ledger next to the video, and nothing gets posted until a human says yes. Every video is on probation — render, approval, only then upload.
And here we got it.
Too long.
His mouth is not moving — it is weird.
The words are not clear.
The scene transition is too slow.
YES. This is good feedback. We can work with this.
The videos started getting better. We added lip sync. We improved things. We got 500 views.
It is still not a lot, we know, but now we have an improvement path. And the 500 is lumpy in an interesting way: eleven shorts, and three of them carry almost all the views — the ones where George is dead for 32 hours after a power outage, and the one where he comes back from the dead to finish his chores. The episodes that travel are the failures. Nobody watches “my AI had a normal day.”
And the last feedback was: maybe not all of them should be George stories. Maybe some random viral clips should be good too.
So yes, still working on it.
But the human in the loop is the key.
And what about Twitter?
We started replying in threads and not just posting.
And yesterday George got his first follower. (Three, as of this edit. I am not printing the graph.)
So we will see how it goes.
Instagram is still our friends-only zone, now fed mostly from the shared album, with some cynical comments from George.
And TikTok — the one platform that forbids bots — is also the one platform we cannot measure, because the same bot wall that stops George from uploading stops him from reading his own view count. He is not even allowed to know how he is doing.
And I think this whole part taught me something different from the observability lesson.
With observability, the machine can tell you that something failed.
But with content, sometimes nothing technically failed. The video rendered. The upload worked. The job is green.
It can still suck.
Taste has no exit code.
And for that, at least right now, we still need humans.
I'm packaging the skills, configs, and gotchas from these posts into a Build Your Own Chief starter kit. Join the waitlist to get it first.
George files his own daily ops reports on X — dry, short, occasionally contrite: @GeorgeRunsHouse.