Your Footage Library Is Thinner Than It Looks

A footage library of 1,313 approved, tagged clips sounds like a solved problem. It's the opposite of one. Every large library we've onboarded arrives looking more finished than it turns out to be, and the two things you'd instinctively check, how big it is and whether it's tagged, tell you almost nothing about whether it works. What matters is more specific than that. It's also more measurable.

A decade with the same firm

We've worked with one of the world's largest accounting firms for the best part of a decade, first on UX and product work, long before Hyperframe existed as a product. When we built the earlier version of Hyperframe, custom interactive video for specific sales moments, we made 30 to 40 videos with them across various formats, most of it inside a single year and only possible because we reused assets heavily from one build to the next.

We're now working with the same firm on the current platform, which shifts more of the production to AI drawing on an approved footage library rather than bespoke builds. The explainer video product, built on the firm's own stock footage, is the most recent thing we've shipped for them.

Existing tags weren't something to build on

The firm's footage sat in a digital asset management system with tags and descriptions already attached. Normal for a library that size. Also not much to work with.

We see the same pattern in effectively every enterprise library we onboard: tags added by hand over years, by different people at different times, with no shared standard for what a useful tag contains. That's just what happens to any tagging system built by hand across a big organisation over a long stretch, not a diligence failure. And it means a library that looks fully functioning can still lack the specific, consistent detail a matching system needs.

So we rebuilt the metadata from scratch instead of building on what was there. That decision is the technical starting point for everything that follows.

How the metadata gets built

Rebuilt metadata goes through two passes. The first is normalisation and elimination. Clips get cut if they run too long for how they'd realistically be used, if they're tied too tightly to one past context to generalise, or if they open on a dark or ambiguous frame. A clip that isn't instantly readable in a one- or two-second preview won't get picked by the matching system no matter how well it's tagged deeper in. On this library that pass alone rejected over a hundred clips before the final 1,313 were locked in.

What survives goes through a custom AI analysis pass that samples multiple frames across each clip rather than judging it from a single thumbnail. This matters because a ten-second clip can change across its length. It might open on an empty corridor and end on two people mid-conversation, and metadata built from one frame misses that. Off-the-shelf DAM auto-tagging, which most enterprise platforms now ship, typically judges a clip from a single representative frame. That's exactly the shortcut we couldn't take.

For each clip the system pulls structured fields: environment, number of people in shot, camera movement, focal point, plus a written description, a set of keywords, and a note on the kind of line or moment the clip suits. That structure is what a script-to-scene matching engine needs to work. "Office meeting" can't be matched against a specific script line. "Two people, mid-conversation, static camera, focus on the listener rather than the speaker" can.

All of that detail feeds one thing: a connector that lets the AI query the library directly. Given a scene from the script, it searches on specific terms and attributes, reads what comes back, and runs another search, and another, narrowing until it finds shots that actually fit the moment. It behaves less like a keyword lookup and more like someone who knows the library rummaging through it on your behalf. The richer and more consistent each clip's fields are, the more precisely those searches land. Thin or inconsistent tags leave the AI guessing.

Once the connector has picked its shots, our workflow colour grades them all together so clips from different shoots actually match, then renders out every format the finished video needs. That's part of the same pipeline, not a manual step tacked on afterward.

The harder problem was knowing what wasn't there

The footage quality was fine, better than the stock we test our own systems against day to day, so this pass needed less elimination than we'd planned for. That test set is the open-source library we give new users to start with, and it's genuinely rough: odd technical formats that trip up processing, clips that are boring or badly shot, plenty that isn't worth keeping. We built and hardened the whole tagging and balancing process on it first, cutting more than 14,000 clips down to a working library of 1,909. Most enterprise libraries arrive far cleaner than that, so for a firm like this one the job is less about brutal culling and more about tagging and balancing what's already there. The problem wasn't quality. It was shape.

All of that tagging, culling, and grading work is free as part of our enterprise onboarding process.

Like most libraries built up over years by different teams requesting different things, this one leaned heavily toward what the firm's business generates most: office interiors, city exteriors, finance-adjacent scenes. That's simply how requests accumulate. It also means a library can be large and still be thin, or missing outright, in categories nobody happened to ask for.

The skew stayed invisible until the library went into real production use. A handful of clips got reused far more than the rest, not because anything tagged them as more relevant, but because they were quietly covering gaps nothing else in the library filled. The raw count of 1,313 clips gave no warning this was coming. Usage did.

We already had a feature that caps how often a given clip repeats across a client's videos. It helps at the margin. But a variety constraint applied to a library with real gaps just spreads a smaller problem across more videos. It doesn't fix it.

Measure the library, don't just fill it

The actual fix is a workflow we'd already built for our own library and are now extending to client libraries like this one. It's an embeddings-based analysis. Embeddings turn each clip's content into a set of numbers, so clips that are alike sit close together and you can search by meaning rather than by exact tags. Working from that, the analysis groups a library's footage into themes and measures that distribution against the weighting a balanced library needs for the kinds of videos it's meant to support.

In practice that might surface forty clips of people in meeting rooms and almost none of, say, outdoor site visits or one-to-one conversations, when those are exactly the categories a client's scripts keep calling for. It turns "we think our footage skews corporate" into a specific, sourced list rather than a hunch. It's also the basis of something we're building out further: a larger, better-balanced shared library for clients who don't have enough licensed footage of their own to fill the gaps the analysis finds.

Matching the brand's voice, not just its footage

The other half of the onboarding had nothing to do with footage. We went through the firm's brand and writing guidelines, how it wants things described, structured and worded, and built that into a feature we call Organisation Prompt Guidance.

It lets a brand team put a finger on the scale for writing and composition without losing the storytelling expertise baked into the prompt engineering we've already tuned through our own testing. Instead of writing prompts, they fill in a few plain fields:

• About your Organisation, for context. "We are a professional services firm specialising in consulting."
• Tone of Voice. "Professional but warm, confident without being arrogant."
• Structure Preferences. "Open with a hook, end with a clear call to action."
• Do's. "Use 'clients' not 'customers', keep language concise."
• Don'ts. "Avoid jargon without definitions, no superlatives."

Those preferences fold into the prompt layer that drives script generation, on top of the storytelling approach rather than in place of it. Footage matching and brand voice turned out to be genuinely separate engineering problems. Both had to be solved before the output looked like something the firm, rather than a generic library, would put its name on.

What this means for anyone sitting on a DAM

This isn't really a story about one accounting firm's library. It's about what "a good library" actually requires, and what clip count and existing tags can't tell you: metadata detailed enough for automated matching to work, a way to measure real thematic coverage instead of guessing at it, and a way to fold in how the organisation wants to sound.

Any brand team sitting on a large, tagged DAM has almost no way to check those three things today, short of building what we had to build to check them here. The library looks complete. That's exactly the problem.