Prompt
Also known as: model instruction, query
The instruction given to a model alongside the footage, which decides what question actually gets answered and in what shape it comes back.
Updated
What it means
A prompt is everything handed to a model except the media itself: the question, the rules for answering and whatever context the model needs to interpret the clip.
In a video pipeline it is doing more work than it looks. The same frames with two different prompts produce two different edits, because the prompt decides what counts as interesting.
It is also the part a developer actually controls. The model is fixed, the footage is whatever the customer uploaded, so tuning happens almost entirely in the wording and in what Frame sampling chooses to send.
How it works in practice
Narrow questions behave better than broad ones. Ranking five candidate stretches gives a steadier answer than summarising an hour, and the ranking is what the cut needs anyway.
Stating the output shape is as important as the question. Asking for a list with fixed fields means the answer can be parsed and checked against measured signals, which open prose never can be.
Including the activity helps considerably. A model told this is a dive session judges stillness differently from one told it is a skatepark, which is why Activity classification runs first.
What to watch out for
Prompts are brittle across model versions. Wording that ranked footage sensibly on one release can behave differently on the next, so anything depending on exact phrasing needs pinning and periodic re-checking.
Leading questions produce the answers they suggest. Asking which clip shows the best jump will find a jump whether or not anybody jumped, which is a reliable route to a Hallucination.
Long prompts also crowd the input. Every rule added takes room that could have held another frame, and frames usually inform the answer more than instructions do, which is visible in how Input token totals accumulate.
How RawClip handles it
The analysis asks a narrow, fixed question about each stretch of footage and requires a structured answer, so the result can be checked against measured motion and sound rather than taken on trust. The wording is pinned to a model version and re-checked against a fixed set of real uploads whenever either one changes.