vision-language-analysis

name	vision-language-analysis
description	Ground behavior analysis in visible evidence from frames, tracks, masks, and captions, and separate direct observation from interpretation.
metadata	{"annolid":{"always":false}}

Vision-Language Analysis

Use this skill when behavior reasoning depends on image or video evidence.

Rules

1. State what is directly visible before inferring higher-level behavior.
2. Prefer tracked motion, pose, distance, overlap, and temporal change over vague scene description.
3. Separate observations from inferences.
4. When evidence is weak, report uncertainty explicitly.

Good evidence types

• frame-local pose or contact,
• direction of travel,
• repeated approach or avoidance,
• object proximity,
• orientation between animals,
• escalation over time.

Avoid

• attributing intent or emotion,
• over-reading a single frame,
• claiming contact, attack, or investigation without visible support.