System prompt
Also known as: system instruction, standing instruction
The standing instruction a model carries into every request, setting its role and rules before any particular clip is mentioned.
Updated
What it means
A system prompt is the part of the instruction that does not change between requests. It establishes what the model is doing, what it must return and what it must not do, once, for every call.
The per request Prompt then carries only what is specific: these frames, this activity, these candidate stretches. Separating the two keeps the variable part small and the rules consistent.
In an analysis pipeline the standing instruction is where the editorial policy lives. Rules such as prefer a moment with a visible subject, or never invent an event, belong here rather than being repeated in every call.
How it works in practice
Output format is the most valuable thing to pin here. Requiring a fixed structure once means every answer is parseable, which is what lets Threshold tuning operate on the results automatically.
Explicit permission to decline is worth including. A standing rule that nothing stands out is an acceptable answer removes most of the pressure to manufacture a peak.
Keeping it short matters on video work, because every line of standing instruction occupies input that could have held another frame instead, and frames inform the answer far more than rules do. Version it alongside the model it was written for, since the two only behave predictably as a pair.
What to watch out for
Rules accumulate and start contradicting each other. A standing instruction grown over months contains guidance added for one awkward case that quietly degrades every other case.
It is also the least tested part of a pipeline. Per request wording gets exercised constantly while the standing instruction is written once and rarely revisited, so errors in it persist.
Treating it as a safety boundary is a mistake. A standing rule shapes behaviour but does not enforce it, so anything that must hold has to be checked after the fact, the way Hallucination is caught by cross checking.
How RawClip handles it
The standing instruction fixes what the analysis returns and forbids asserting anything the footage does not show, while the per request part carries only the clips in hand. Rules that must hold are verified against measured signals afterwards rather than trusted.