# How to pull product shots from client footage in Claude

> Describe the shot, do not describe the file.

Canonical URL: https://vivu.ai/guide/how-to-pull-product-shots-from-client-footage-in

Describe the shot, do not describe the file. Claude cannot see inside a video by itself, so the working setup is a video search connector holding the client's footage, and a request phrased the way you would brief an assistant editor: the product sitting on a desk with the label facing camera, or a hand picking the product up off a shelf. What comes back is a set of time ranges you open and check before anything gets exported.

The phrasing matters more than people expect. Describing what is visible in the frame works. Describing what the shot is for does not, because nothing in the footage marks a clip as the hero shot.

## The available routes

**Scrubbing the timeline.** Still the fastest option for a single clip you half remember, and it does not survive a client handing over several hours of raw material from a shoot you were not on.

**Logging on ingest.** Someone watches everything once and writes down what is in it. This produces the best search results of any method here and costs a person's day per shoot, which is why agencies stop doing it in month three.

**Transcript search.** Cheap to set up and blind to anything silent. A product shot is usually silent, so this route answers almost none of the questions you have about client footage. [Automatic tagging](https://vivu.ai/guide/what-software-can-automatically-tag-video-files) is the adjacent version of the same trade: it generates labels, and you get whatever vocabulary it happens to use.

**A video search connector on the footage.** The client's material is uploaded and indexed once, then queried by description from inside a chat.

## Pull a product shot out of a client's long footage

[Vivu](https://vivu.ai/mcp) is added to Claude as a custom connector, after which it appears as a tool the chat can call. One client gets one project, and you ask for the shot by what is in frame: `a wide shot of the product on a table with the window behind it`, or `someone holding the product and turning it toward camera`. Composition works well as a query, because framing and background are things the footage actually shows. So does posture and action, which is how you find the same person appearing at several points across a long recording without knowing their name. Each result comes back as a time range with a line of reasoning describing what is in it, the results page previews them one at a time, and for the ones you keep you can export the original clip to hand to the editor.

Write the query with the elements you care about and the reasoning will check them off one by one. Ask for a wide shot of the product on a table with a window behind it and you get told which of those it matched, which is faster to scan than the footage is.

## Where this route stops

The reasoning describes appearance, and it will not tell you who someone is. If the client needs shots of a specific person, you have to put the visible features into the query yourself, which is fragile when three people at the shoot wore the same shirt. An empty result is likewise ambiguous. Ask for a tight close-up of the product and get nothing back, and you cannot tell from the result whether the shoot never captured one or whether your description was stricter than the footage. Loosening one element and asking again usually resolves it.

The structural limits matter too. Each search runs inside one project, so this is one client at a time by design. The footage has to be uploaded to a Vivu project and indexed in the cloud before you can query it, which is a step, not a background process that happens to material sitting on a drive. Searching consumes an allowance. And a time range is something you open, not a frame-accurate mark, so the editor still trims.

Also worth saying plainly to clients: this finds the shot, and none of it cuts anything. The export hands the editor a clip and the edit is still an edit.

## When you should skip it

A single shoot with an hour of material and a shot list you wrote yourself does not need any of this. Neither does footage you are about to cut this week and never open again, since indexing pays off across the second and third time you go back to the same material. The case where it clearly earns its place is a client relationship with a growing pile of past shoots, where the brief for this campaign keeps referencing footage from the last one.

## Deciding

Ask how often you go back into old client footage. If the answer is rarely, timeline scrubbing is correctly priced and the setup cost is not worth carrying. If the answer is constantly, and if the person who was at the original shoot has left or is on another account, then the real work is turning a description into a location, which is the thing this route actually does. [How the search itself works](https://vivu.ai/guide/how-does-ai-search-inside-video-footage) is the piece to read if you want to understand why descriptions of visible things work and descriptions of intent do not.

## FAQ

### Can I ask for a shot of a specific person by name?

Not reliably. The search matches descriptions of what is visible, so it has no notion of who anyone is unless you tell it what they look like. What works is describing observable features and behaviour: the person in the dark jacket standing at the left of frame, or the person holding the microphone and gesturing. That gets you the right shots when only one person at the shoot matches the description, and it gets confused when several do. For a shoot where the client needs a named individual, a written shot log from the day remains the more dependable index.

### Does exporting a clip mean the tool edits my footage?

No. Exporting hands you the original piece of video for the range you selected, and nothing is cut, re-encoded into a finished sequence, or generated. The file goes to whoever is doing the edit and the edit happens in your editing software as usual. Anything that promises a finished cut is doing a different job than finding footage.

### What happens if the client's footage is in an unusual format?

Common video formats upload directly. The exception worth planning around is anything that is not really a video file, like a log or data container from another kind of recording system, which has to be exported to video first. If a client hands over material from a capture setup rather than a camera, check what the files actually are before you budget time for the ingest.

### Can I search across all my clients at once to find a reference?

No. Searches are scoped to a single project, and since each client gets their own project, a query reaches only that client's material. For reference hunting across accounts you are back to your own notes. This limit is often a feature in an agency setting, because it also means a query about one client cannot surface another client's unreleased footage.