review_generation checks the prompt, not whether the clip will go viral

review_generation checks the prompt. It returns whether the image or video matches, a list of issues, a severity of ok / minor / broken, and a suggested fix. It is not a virality score. It is not brand-kit hex compliance. It is not a second generate. Skip it, reroll, and you have bought another roll of the same unread failure. Run analyze_trend on the file instead and you have bought a research report on a question nobody asked.

Get AI feedback on a generation is the named agent job. Point at a generation (or say "latest") and, for a prompt-match check, the original prompt. Required: media_url, media_type, original_prompt. It is a lightweight vision check, not a new catalog generate. Versely does not run it automatically after every generation because of extra cost and latency. Ask when you want it — after a long workflow or movie, or when you are unhappy.

For open-ended composition notes rather than a strict match, supporting analyze_content can critique quality. That is still not "will this hit the For You page."

What the job is

A structured verdict on this output versus this brief. A "minor" with a named region is a salvage signal: extract a frame, edit a still, do not reroll the universe. A "broken" with a suggested fix is the next prompt, not a vibe.

Offer it after a long job. Never treat silence as a pass. Silence was a cost decision, not a quality decision.

What it is not

• Trend analysis. Analyze a viral video or post needs a public social URL. Your generate is not that URL.
• A full beat sheet. Get a detailed breakdown of a video extracts frames, reads on-screen text, optionally transcribes. Use it to understand a reference. Use review to QC your take against your prompt.
• Brand-manual audit. The kit's hex is not the original_prompt. Review will not fail a near-red that isn't #FF5733 unless you wrote the hex into the prompt it is checking.
• Edit or remake. Edit or redo a previous generation spends a model. Review spends a check. Check first when you cannot name the fault.

The cheap-looking path is "just generate again." That is a full model price to avoid a vision pass. On video, that is the wrong order.

The test

Do you have an output, the prompt that was supposed to make it, and a question of the form "did we get what we asked?"

Yes: this job. Run the verdict. Then extract, edit, or remake according to severity.

If you wanted "why did their Reel work," analyze_trend. If you wanted "what happens in this file," analyze_video. If you wanted a new clip, you still may — after the verdict, not instead of it.

FAQ

Why isn't this automatic on every generate?

Cost and latency. A vision pass on every still would tax the small jobs. Ask on the ones that are expensive to reroll, or when something feels off.

Can I use it without the original prompt?

The match tool requires original_prompt. A general critique can go through analyze_content. If you do not remember the prompt, pull it from history rather than inventing a new brief to "check."

Does a "broken" verdict refund the generate?

No. It tells you not to ship it and how to change the next confirm. The generate already spent. That is why the check exists: to stop the second spend from being blind.

Will it tell me which model to use next?

It may suggest a fix in prompt space. Switching families is still your (or the agent's) routing decision. Review is QC, not a model catalog.