Agentic Video Understanding turns a retrieval goal into a short list of moments from a long recording. You supply a source (a URL or local media path) and a goal such as “the exact next-step commitment and any pricing pushback”. The skill restates the goal as one to three testable queries, calls Gemini’s agentic video understanding, and keeps only the moments that carry real timestamps.
Each kept moment has a start and end time, the modality it came from (frames, audio or transcript), a quote, a flag for verbatim or paraphrase, a reason and a confidence score. The skill stops there. It never cuts, captions, schedules, posts or writes to a CRM, so the calling workflow keeps every approval.
What You Get
- A list of up to five moments by default, each with start and end time, modality, quote, reason and confidence
- A verbatim or paraphrase flag on every quote
- An explicit empty result with a reason when nothing matches
- A note on whether the agentic path or a fallback static ingest was used
- Output as Markdown for people and an optional JSON block for other tools
Common Use Cases
- Mine a sales call for the buyer’s objection, next step or competitor mention before the follow-up
- Find a three-second standalone hook and in and out points for a short clip
- Pull named-person and contrast moments from podcast or YouTube tape for a long-form post
- Review a Loom or trial recording for evidence that someone meets a role bar
- Audit a discovery recording for every mention of a keyword such as budget
What Makes It Unique
Most video skills cut, caption or publish. This one only retrieves evidence and then stops, with hard gates against invented timestamps and quotes, against full fixed-frame-rate ingest when the agentic path exists, and against any send, post or CRM write. That makes it a safe retrieval layer that other video and sales skills can call.
Under The Hood
It is a prompt-only skill with no scripts. The agent restates your goal as queries, then calls the Gemini API (or Google AI Studio) with the media source, which needs a Gemini API key in the environment. The model reads the video or audio remotely; the skill itself writes no files and runs no commands. It normalizes the response into the output schema and drops any moment without a real timestamp.