1. Three priorities, one list
- MUST: a viewer will notice. Fix before release.
- SHOULD: small but visible. Fix it.
- CHECK: needs eyes or ears. The judge or the owner decides.
Each item has a time, the shot, what is wrong, and a suggested fix. Before flagging anything, check the script: a shadow walking past a bakery looked like a stranger, but the script had asked for it.
2. Cheap fixes first
Many fixes need no new video: swap a dissolve for a hard cut, lay the ambience bed, trim before a prop vanishes, move a title down. Those are almost free. The rest are a remade still, checked, then a new clip. If every take of a shot fails, fix the still rather than the clip; if that fails, simplify the shot.
In this project the first fix round cleared every MUST in three episodes in one afternoon. A second, automatic slop review found nine more problems the first review missed, and the maker cleared those within hours.
3. Version, log, re-check only what changed
Every fix round is a new file next to the old one, never an overwrite, with a change log: shot, what changed, time in the new version. The reviewer re-runs the checks and looks at exactly the changed shots. Everything else was already passed.
4. What still needs a person
“Zero slop with no human” is not possible today. “Near-zero slop with a few short human gates” is. The gates are:
- Approve the reference pack once.
- Glance at the stills before they are animated.
- Clear the short queue where the checks disagree.
- Watch each finished episode once, with sound: voices, emotion, lip sync, music, and whether the story makes sense to a child.
That watch also feeds the escape log, so every human catch makes the pipeline permanently a little stronger.
5. Music, last and separately
Keep music out of generated clips. Build a small licensed library of short instrumental cues (theme variations, moods, stings) with a music generator whose plan allows commercial use, tag each scene with a mood, and let the edit rotate cues so episodes do not all sound alike. Lower the music automatically under speech, then run the loudness checks again.