# How to turn client ad footage into briefs with Claude

> Claude writes the brief. It does not watch the footage.

Canonical URL: https://vivu.ai/guide/how-to-turn-client-ad-footage-into-briefs-with

Claude writes the brief. It does not watch the footage. That splits the job in two, and only the first half is hard: getting the specific moments out of a client's spots in a form an assistant can read, then handing those over to be written up. Once you have the moments, any capable assistant will turn them into a teardown or a shot list. Until you have them, prompting harder does nothing, because the model has no way to see what is inside the file.

## What a brief needs out of the footage

A brief that is any use to the person making the next ad points at positions. Where the hook lands. The first frame the product is actually on screen. The moment the offer gets said out loud. The frame where the price card comes up. None of that lives in the file name, and most of it does not live in the transcript either, because a price card and a product shot make no sound. So the real question is how a moment described in words becomes a timecode somebody can open.

## The routes people take

Attaching the spot to the conversation each time works, and it works for exactly one short file. You are still the one who found the moment, you redo the whole thing in the next conversation, and a folder of a client's quarterly output is not something you push into a chat window.

Pointing an assistant at the file system gets you an assistant that can read names, folders, and dates. It can rename, sort, and reason about structure, which is genuinely useful and completely blind to content. A file called `hero_v4_final.mp4` tells it nothing about what is in frame at eleven seconds. [What an assistant can decide from names alone](https://vivu.ai/guide/using-claude-to-organize-files) has a fairly hard edge, and this is where it sits.

Building the index yourself means transcription, embeddings, a store, and a way to expose it. For ad footage this is where transcript-only pipelines quietly fail: the beats that matter in a thirty second spot are visual, and a transcript index returns nothing for them. [Searching by description rather than by words spoken](https://vivu.ai/guide/is-there-a-tool-to-search-video-by-text) is a different kind of index, and it is the kind this job needs.

The fourth route is a hosted retrieval layer the assistant can query directly, so the finding happens where the writing happens.

## Pull the beats out of a client's spot, then have Claude write them up

[Vivu](https://vivu.ai/mcp) can be added to Claude as a custom connector. With that client's spots uploaded into that client's project, you ask for beats the way you would describe them to an editor: the moment a price or cost breakdown appears on screen, or the first shot where someone holds the product. What comes back is a set of openable time ranges, each with a line explaining why it matched, and the result page lets you preview them one at a time before you trust any of them. From there Claude is the one doing the writing, arranging the ranges you confirmed into a brief or a teardown table, since the retrieval side finds footage and does not cut or generate anything.

Two things about the output are worth knowing before you build a process on it. On short spots, a broad first pass mostly tells you which files are relevant rather than which seconds, so getting down to a beat means running the slower precise pass. And when a match reason quotes text that was on screen, that quote is there so you can scan the list, not so you can paste a number into a brief. Open the range and read it off the original.

If what you are tearing down is a competitor's ad rather than your client's own footage, check that you have the right to use the material before it goes into a project.

## When this is more setup than the job deserves

One spot that you shot last month, with beats you already remember, does not need any of this. Neither does a brief that lives at the positioning level and never cites a moment. The setup starts paying when you are working across a client's back catalogue, or when the person writing the brief is not the person who was on set.

## Where this route stops

Each search runs inside a single project, so there is no question you can ask once and have it sweep every client you work with. Footage has to be uploaded and indexed in the cloud before any of it is searchable, which means the first day with a new client is an upload day. Searching draws down an allowance, so a habit of firing off vague queries costs more than thinking about the query first. And what comes back is a range you open, not a frame-accurate in and out point, so the editor still makes the cut.

## Deciding whether it is worth it

The test is how often you are asked to explain footage you did not personally shoot. If briefs mostly get written by whoever was on the shoot, from memory, the finding problem is not yours yet. If they get written weeks later from a folder of exports, by someone reconstructing what happened, then the bottleneck is never the writing, and a faster assistant on the writing half will not move it.

## FAQ

### Can Claude read a video file I attach to the conversation?

An assistant works from whatever you hand it, and handing over one short spot is workable. It stops being a workflow as soon as there is more than one file or more than one conversation, because you re-attach every time and you are still the one deciding which part of the footage matters. The thing that scales is a library the assistant can query, where you describe the moment and it comes back with where to look.

### Can I search across all my clients' footage at once?

No. Each search runs inside one project, and the practical setup is one project per client. That is a real limit if you were hoping to ask which of your clients has ever shot a particular kind of scene, and it is the behaviour you want when the material is somebody else's property.

### What if the ad has no dialogue at all?

Then describe what is on screen instead of what is said. Queries like a price appearing on screen, or a close shot of a hand on a device, work on the picture rather than the transcript. Ads are the case where this matters most, because a large share of the beats in a thirty second spot are visual and a transcript-based search has nothing to match against.

### Does the brief come back ready to send to a client?

No, and you should not want it to. What the retrieval step gives you is a set of candidate ranges with a reason attached to each, several of which will be near misses that only a person can rule out. The assistant can draft around the ones you keep, and the judgment about which moments are the argument stays with you.