# How to write a video brief from footage in Claude

> Write the brief from shots you can open, not from memory of what the last round looked like.

Canonical URL: https://vivu.ai/guide/how-to-write-a-video-brief-from-footage-in

Write the brief from shots you can open, not from memory of what the last round looked like. The workable version is a video search connector holding your existing ads and product videos, a conversation where you ask for the beats you want the next round to hit, and a brief that Claude assembles out of the ranges you kept. The assistant does the writing. It cannot watch anything, so the finding has to come from a tool that can.

This matters because briefs written from memory drift. Someone remembers the opening of the spring campaign as tighter than it was, and the new brief specifies something nobody actually shot.

## Ways teams build a brief from existing work

**From memory and a mood board.** Fast, and it encodes whatever the loudest person in the room remembers. Usable when the team is small enough that everyone saw everything.

**Re-watching last quarter's cuts.** Accurate and expensive. Someone spends half a day and comes out with notes that are correct and already out of date by the next round.

**A spreadsheet of shot notes.** The version that works if you maintain it, which is the catch. Teams shipping several product videos a week stop updating it within a month, and a partly maintained log is worse than none because you trust it.

**Searching the footage by description.** The existing ads and product videos get indexed once, then you ask for the beats directly and get back ranges to open.

## Build the brief from shots you can open

[Vivu](https://vivu.ai/mcp) connects to Claude as a custom connector. The queries that produce a usable brief are the ones written as beats rather than as judgments: `the first frame where the product appears on screen`, `the moment a price or cost breakdown is shown`, `the part where the main benefit is said out loud`. Each result comes back as a time range with a line of reasoning describing what it found, and the reasoning will often restate what was on screen, which is enough to tell at a glance whether it is the beat you meant. From there Claude can lay out what you kept into the structure your team briefs in, with the ranges attached so the next person can open them.

There are two search modes. The precise one narrows to the seconds where a beat actually happens and runs as a background job, so you ask, do something else, and collect the results when it finishes. That wait is worth building into how you work rather than treating it as a hang.

## The part where you still have to check

The reasoning attached to a result is a description of the footage, not data extracted from it. When a result says a cost breakdown appeared on screen and quotes the figures, those figures are how the on-screen text was read, and they do not belong in a brief until someone has opened the range and confirmed them. This is especially live with concept pieces and placeholder mockups, where the numbers on screen were never real to begin with and will read as real once they are sitting in a brief.

The route also has fixed edges. A search runs inside one project, so this is one body of work at a time. The videos have to be uploaded to a Vivu project and indexed in the cloud before you can ask anything about them. Searching draws on an allowance, which is a reason to write fewer, better queries. And the results are ranges to open, not frame-accurate marks, so nothing here replaces the timeline.

The larger limit: a brief built this way describes what exists in your footage. It cannot tell you which of it worked, because performance data lives in the ad platform and is not visible to a search running over video. Anyone promising a brief that knows which hook won is describing a different tool.

## When you do not need this

A team producing a few videos a quarter, all of them shot by the same two people, can brief from memory and be right. A brand refresh that deliberately breaks from everything shot before does not benefit from an inventory of what came before. The case where it pays is the one where the same product gets shot again every few weeks, versions accumulate, and nobody can say with confidence what the last four rounds actually did. Keeping that legible across handoffs is the same problem [a workflow that survives handoffs](https://vivu.ai/guide/content-marketing-workflow-for-video) is built around.

## Deciding

The test is whether your team can currently answer "what did we do last time" without opening files. If yes, the brief you write from memory is fine and setting this up is overhead. If the answer takes a day of digging, and if that day happens before every round, then the work is not writing the brief. It is reconstructing what exists, and that is the part worth moving. Before committing to anything, it is worth checking whether your problem is actually retrieval at all, since [footage you cannot locate](https://vivu.ai/guide/how-to-find-old-marketing-videos) and footage you can locate but never re-examine are different failures with different fixes.

## FAQ

### Can Claude write a creative brief by watching our old ads?

No, not on its own. An assistant works with text and cannot open a video file and see what happens in it. What it can do is take descriptions of moments that a video search tool returned and assemble them into a brief. The distinction shows up the first time someone asks an assistant to analyze a campaign it has no access to: the output will be confident, well structured, and made up.

### Will this tell me which of our ads performed best?

No. Searching inside video sees what is in the footage and has no view into what happened after the ad shipped. Impressions, hold rates and conversion all live in the ad platform, and nothing about indexing your videos brings that data into the picture. What you get is an accurate inventory of what you made, which is a genuinely useful input to a performance conversation and is not the conversation itself.

### How specific should a query be when looking for a beat?

Specific about what is visible or audible, vague about intent. A request for the moment a product first appears on screen works because that is an observable event. A request for the most compelling moment does not, because nothing in the footage marks compellingness. When a request comes back empty, the usual cause is a description with too many conditions stacked together, and dropping one of them is more productive than rewording the whole thing.

### Can it pull the on-screen text out of our videos as data?

Treat anything it reports about on-screen text as a pointer, not as extracted data. A result may describe what a caption or price card said, and that description is how the footage was read rather than a verified transcription. It is reliable enough to tell you which range to open and not reliable enough to paste into a brief unchecked. Anything that will be quoted, especially numbers, gets confirmed against the original.