1. Slop is anything a viewer notices that nobody meant
“AI slop” is a loose word, so this series uses a tight one. Slop is any moment in a finished video that a viewer would notice and that nobody wrote. A loaf floating in the air when the script says it is under someone’s arm. A jumper that becomes a coat between two shots. A sign whose letters turn to nonsense. A line of dialogue that is simply missing. Two seconds of dead silence in a busy street.
Slop is different from a style choice. A soft, painterly look is a choice. A child with six fingers is slop. The test is simple: would the person who wrote the script say “that is not what I asked for”?
2. The first attempt: a month of credits, no episode
The project was a short animated series for young children, made with an image-to-video model run by AI agents. The first month produced plenty of footage and no episode anyone could show. When the team asked the production agent for its own account of what went wrong, the answer ranked the causes:
- Too much at once. One episode was pushed through as 89 shots in a single run: 419 image calls and 174 video calls. The same defects repeated across the whole cut, because nothing was proven small first.
- Polishing the wrong thing. About 420 image calls went on making still frames perfect. The real failures, hands, floating and morphing, only show up once the still starts to move.
- Checking by numbers. The quality checks scored speech-to-text, pitch and face similarity. They passed shots with six fingers, a floating child and a plate that turned into a letter.
- Three bosses. Three agents gave production orders with different rules. A decision would be reversed within minutes, and nobody could tell which instruction was the owner’s.
- Scripts built on the model’s weak spots. Small hands counting props, children sitting on furniture, food that had to hold a shape, animals acting, group singing. And the story kept changing while it was being made.
- No brakes. A standing instruction not to “compromise on calls” meant no retake limit.
3. The one thing that worked
In the middle of all that, one small test was approved by the owner: three clips, three requests, about 0.1% of the weekly allowance each. Every one of those clips had the same shape. One child in frame. One line. A held pose. No prop in the hands. Four to six seconds. A detailed prompt with the exact line in quotes and a fixed description of the voice.
That proof became the template for everything in this series. It is the same lesson every production learns: make the smallest thing that can be judged, judge it, then repeat it.
4. What changed afterwards
The rest of this series follows the rebuilt process in order:
- Design shots the model can actually make.
- One maker, one reviewer, one approver.
- Lock pictures, not adjectives: a reference pack.
- Assemble without creating new slop.
- A scan that reads every frame.
- Read every word on screen.
- Check every spoken line against the script.
- A judge that asks narrow questions.
- The fix loop, and the one human watch that remains.
The checking tools run on one Mac, for free, and take about two to four minutes per episode. None of it removes the human entirely. Lesson 10 says honestly where a person is still needed.