Claude cannot watch your recordings on its own, so the working setup is a video search connector holding your webinars, customer interviews and event footage, and a request made while you are drafting: find the part where someone explains why they switched. You get back time ranges you can open, you pick one, and it goes into the post. The assistant does the asking and the drafting around it. The footage gets searched by something built to search footage.
Most marketing teams already have the b-roll. It is sitting in a year of webinar recordings that nobody reopens because reopening them means scrubbing.
The routes people actually take
Stock libraries. Fast, safe, and the reason so many B2B posts are illustrated with a stranger pointing at a whiteboard. Fine for atmosphere. Useless when the post is about your product.
Asking the person who was there. Surprisingly effective and completely unscalable. It works until the person who ran that webinar changes jobs, at which point the footage becomes unsearchable in practice even though it is sitting right there.
Transcript search. Export the transcript, use find, note the timecode. This is genuinely good for spoken content and is the route most teams should try first, since it costs nothing and catches the majority of quotable moments. It goes blind the moment you want something visual.
A video search connector in the chat. The recordings get uploaded and indexed once, then a request in the conversation returns openable ranges without leaving the draft.
Pull a clip for the post while you are still drafting it
Adding Vivu to Claude as a custom connector is a one-time step. After that, the useful habit is asking for the moment by its shape rather than its wording: a short direct instruction telling the audience what to do, or the part where a customer describes the problem they had before switching. You do not need to remember the exact phrasing, which is the difference from transcript search, where you have to guess the words. What comes back is a set of time ranges with a line of reasoning for each one, previewable on the results page, and because you are already in the conversation where the post is being drafted, the range you pick and the paragraph it illustrates get written next to each other instead of in two separate sessions a day apart.
One practical note from using this on recorded panels: the ranges that come back can be very short, a few seconds around the thing you asked for. Before a clip goes into a post, extend it at both ends, since a clip that starts mid-sentence reads as sloppy even when the content is exactly right.
Where it breaks
A request can also land on the setup rather than the payoff. Ask for a direct instruction to the audience and you may get the moment someone says "here is what you should do," with the actual instruction arriving after the range ends. You catch this by previewing, which is why previewing is not optional.
Empty results are the other thing to expect. If nothing comes back, that does not confirm the moment is missing. It may mean the description was narrower than what the recording contains. Drop one element of the request and ask again before concluding the footage does not have it.
The structural limits are worth knowing before you plan around this. Each search runs inside a single project, so a query cannot sweep your whole video history at once. Recordings have to be uploaded to a Vivu project and indexed in the cloud first, which is a real step for an archive that currently lives across a drive and three meeting platforms, and it is not something that happens automatically to new recordings. Searching draws on an allowance. And what you get is a range to open rather than a frame-accurate cut point.
When this is not worth setting up
If your blog publishes monthly and each post gets one image, stock is the correct answer and always will be. If your recordings live in a meeting platform your team has not yet decided to keep, sort that out first, because there is nothing to index. The threshold where this changes is usually when someone starts saying "we definitely said that on a webinar" more than once a quarter. Below that, the plain version of finding an old video covers it without new tools.
Deciding
The question to settle is whether your bottleneck is finding footage or having footage. Teams that have not recorded much should not be shopping for search, and teams sitting on two years of webinars usually already know which posts they would illustrate differently if the clip were one question away. If your archive is mostly spoken content and you are comfortable with transcripts, start there and see how often it fails you. The moments it misses are the visual ones, and counting them over a month is a cheaper way to decide than any pitch. What an assistant can and cannot work out from files it cannot watch is laid out in what Claude can decide about files, and it is worth reading before you assume the assistant is doing the looking.
FAQ
Can Claude search my Zoom or webinar recordings directly?
Not by itself. An assistant can read a transcript you give it and can call tools that have been connected to it, but it has no independent access to a recording sitting in a meeting platform and cannot watch video in any case. For a recording to be searchable by content, it has to be in a system that has indexed it. Exporting from the meeting platform is a step you take, not one that happens in the background.
Is transcript search good enough for finding b-roll?
Often, yes, and it is the honest first thing to try because it costs nothing. If the clips you want are people saying things, a transcript plus your editor's find function will locate most of them. Where it fails is anything without words: a product on screen, a room reaction, a demo happening while someone talks about something else. It also fails when you remember the idea but not the phrasing, since find matches the words you type rather than the meaning.
How long are the clips that come back?
They vary with what you asked for. A request for a single spoken line tends to return something very short, sometimes only the few seconds around it, while a request about a longer stretch of discussion returns something correspondingly longer. Plan on extending whatever comes back before it goes into a post. The range marks where the thing is, and the framing of the clip is your call.
What if the clip I want is in a video nobody uploaded?
Then it is not findable, and no amount of query phrasing fixes it. This is the unglamorous part of any search setup: coverage is determined by what got ingested. Teams tend to discover this when the first few searches come back empty and the cause turns out to be an archive that is only half in the system. Deciding upfront which recordings are worth keeping, and which ones nobody will ever reopen, is worth more than tuning the queries.