A publish review is the last credit, not a substitute for picking a take
analyze_video and review_generation cost credits and time. The tools themselves say not to run them on every generate. They grade a file that is about to ship.
The last step before a paid placement is watching the file. analyze_video extracts frames, beats, on-screen text. review_generation grades a generate against its original prompt and returns matches_prompt, issues, severity (ok / minor / broken), and a suggested fix. analyze_content sits nearby. Both reviews cost credits and latency. review_generation's own guidance is not to run it on every generation. Reserve them for files about to publish or about to be promoted. That is the definition of a later step.
Check a video before you publish it is the job. estimate_cost before a deep pass on a long video. This is a review, not a repair. The fix is a separate edit.
QC is not how you choose a take
People run review_generation on take one because it feels like a director. It is a vision pass against a prompt. If you already know the eyes are wrong, you do not need a structured verdict. Regen. If you are about to spend distribution on the keeper, then pay for the watch-back: misspelled overlay, caption in the chrome, a prompt mismatch you stopped seeing.
review_generation needs the original prompt. It is for AI generations, not an arbitrary camera original. Uploaded footage is analyze_video. analyze_video samples frames; a single-frame glitch can slip a shallow pass. Raise depth when you are hunting that, on a file that matters.
Do not auto-chain the review after every generate in a taste loop. That is the leak the tool text warns about.
The AI video editor is often where the fix happens. Review points; it does not patch.
What has to be true
- You have already picked this file over the bin.
- Picture, mix, and type are the ones you think you will ship.
- A broken caption would now be more expensive than the check.
Then run it. Then fix. Then export. Then publish. The review is not a first step, a middle coping mechanism, or a substitute for looking with your own eyes at 1×.
FAQ
Can I use review_generation as a cheap taste test?
It is not cheap, and it is not taste. Taste is you, watching. The structured verdict is for a generate you have already almost accepted.
Does a "broken" verdict regenerate the shot?
No. It suggests a fix. You still decide whether to regen, patch, or ignore.
Should I review before captions?
Review the file you will publish. If captions are part of that file, burn them on the locked cut, then check. Checking a naked miss and captioning later means you never checked the type.
Why does analyze_video miss a one-frame flash?
It samples. Deepen the pass on the keeper if you are hunting flashes. Sampling every generate at max depth is the leak.