Under the hood

How the agent decides which clip to show

An AI casting b-roll cannot watch your footage. It reads what you wrote about it — which is why the Asset Library is a described library, not a folder.

A folder is not a library

Point an automated video tool at a folder of MP4s and it has almost nothing to work with. Filenames and durations do not tell it that one clip shows a timeline being scrubbed and another shows a settings panel. So it either picks at random, or it does not pick at all.

Orb Studio's Asset Library stores two written fields against every asset, and they are the whole mechanism. What's happening in this asset describes what is visibly on screen. When should the agent use it is casting guidance, written as an instruction — the field the art director actually reads when it is deciding what to cut to.

Write the usage field as a rule, not a summary

The common mistake is to write the second field as a restatement of the first. “A clip of the timeline editor” tells the agent nothing it did not already know. “Use when the script claims the output stays editable rather than being a locked render” tells it exactly which sentence this footage belongs under.

Good usage notes carry their limits too. If a clip shows an empty state, say so, and say not to cast it when the narration is claiming completeness. The agent will follow that. A library where every asset claims to be perfect for everything is a library the agent cannot reason over.

Tags, availability, and what happens when a file moves

Tags narrow the search — Product Demo, Intro, Outro, Music, Voice, and anything custom you add. A per-asset Available to agents toggle decides whether automated video can use it at all, which is how footage that is useful to you but wrong for public output stays in the library without ever being cast.

The library is also missing-media aware. Move or delete an asset's file outside the app and the card flags itself, the Editor marks its clips offline, exports refuse to start until it is resolved, and the agent stops casting it. Put the file back — or relink it — and every project using it plays again, with name, description, tags and agent settings intact.

The shortest path to a described library

Writing all this by hand for a hundred clips is the reason most libraries stay undescribed. Record mode has a shortcut: mark sections as you record, talk through what each one shows, and at Mark Out the app transcribes your narration on-device and drafts the name, description, when-to-use and tags for one-tap approval. Every approved section is cut from the take and lands in the library, agent-ready. One recording session becomes a described shelf.