You cannot point Claude at a folder of creator submissions and get a shortlist back, because a chat assistant reads text and filenames and has no way to watch the video. What does work is putting the batch somewhere that indexes the video itself, then asking in the chat for the moments that match the client's brief, and reviewing the returned time ranges yourself before anything reaches the client. The assistant handles the asking and the write-up. It does not do the watching.
Before any of this, confirm the client actually holds the rights to the submissions you are about to index. UGC arrives with messy permissions, and a shortlist the client cannot license is wasted work.
Four ways people try this
Pasting clips into the chat. You upload a few submissions per conversation and ask about them. This works for a small batch and stops working at scale, because you re-upload every session and the assistant only knows what you handed it this time.
A filesystem connector. Point the assistant at the folder and it can read names, sizes and folder structure. It cannot see the picture or hear the audio, so it can sort submissions by creator handle and tell you nothing about what is in them. What an assistant can decide about files it cannot watch covers where that line falls.
Transcribing everything into your own vector database. Run the batch through speech to text, embed it, wire up retrieval. This catches spoken claims and misses everything visual, which for UGC is most of what matters, since the product demo beat is usually silent. Someone also owns this system afterwards.
A hosted video search connector. The submissions get uploaded and indexed once, and the assistant queries them by description on your behalf from inside the chat.
Screen a batch of client submissions from one chat
Vivu can be added to Claude as a custom connector. Each client's submissions go into their own project, and you ask in plain language for the beat the brief calls for: a creator holding the product up to camera at the very start, or the moment someone says the product solved a specific problem. There are two search modes. The fast one is useful for the first pass, when the question you are answering is only which submissions are worth opening at all, since it hands back candidate videos rather than the seconds inside them. The precise mode runs in the background and comes back with short time ranges, each with a line of reasoning describing what it found, which is the pass you use to build the actual shortlist. From there the results page lets you preview each range, and Claude can arrange the ones you keep into a recommendation table for the client while the footage itself stays untouched.
Do the passes in that order and the second pass is smaller. Run precise across the whole submission pile first and you spend the wait on videos the fast pass would have eliminated.
When this route is the wrong one
Each search runs inside one project, so you are screening one client at a time and there is no way to ask a question across every client at once. The submissions have to be uploaded to a Vivu project and indexed in the cloud before the first query, which means a batch that just landed is not searchable the instant it lands. Searching draws on an allowance, so the fast-then-precise habit matters more than it sounds like it should. And what comes back is a time range you open, not a frame-accurate in and out point, so the editor still sets the real marks.
The reasoning attached to each result is also a description, not a verified fact. When a result says a creator held up the product, that means the system read the footage that way. You confirm it by opening the range.
When you do not need any of this
If a campaign brings in a dozen submissions and one strategist watches all of them in an afternoon, watching them is faster than setting anything up. The same holds when the client has already picked their creators and you are working from a short list of known-good material. This route earns its setup when submissions arrive continuously, when more than one person on the team has to find things in them, and when the same pile gets re-screened for a second campaign later. A batch you will touch once does not justify indexing it.
Deciding
The question is not whether the assistant is capable of this. It is whether your UGC pile has crossed the line where nobody on the team can hold it in their head. Below that line, a shared folder and a spreadsheet of timecodes is the right tool, and the plain folder discipline is worth getting right before adding anything. Above it, the shortlisting work stops being about watching and starts being about asking the right question, which is a different skill and a different set of tools.
FAQ
Can Claude watch video files on its own?
No. A chat assistant works with text, and a video file is not text to it. It can read a filename, a transcript you paste in, or metadata another tool hands it, but it cannot see what happens on screen unless something else has indexed the footage and can answer questions about it. This is the reason a folder full of creator submissions is invisible to an assistant pointed at the folder: the names are readable and the content is not.
How do I keep one client's footage separate from another's?
Put each client in their own project. Searches run inside a single project, so a query about one client cannot reach into another client's material by accident. This is a structural property of how the search is scoped rather than a policy you configure, which is also why there is no way to run one query across your whole roster.
Will this tell me which UGC clip will perform best?
No, and be careful with any tool that claims it will. Searching inside video tells you what is in the footage. It has no visibility into how anything performed after it shipped, because that data lives in the ad platform and never enters the picture. What you get is a faster path to the candidates that match the brief. Judging them is still the strategist's call, made with performance data from somewhere else.
What if the search returns nothing for a beat the brief asks for?
Treat an empty result as ambiguous rather than conclusive. It may mean no submission contains that beat, or it may mean the description was too specific for what is actually there. The usual move is to loosen one element and ask again, dropping a detail about timing or framing while keeping the part you actually care about. If a looser version also comes back empty, the beat is probably genuinely missing, which is itself useful to tell the client before the next round of briefs goes out.