Agentic video understanding
Use when an agent must extract moments, quotes, objections, hooks, or evidence from long video or audio cheaper than full-frame ingest — sales calls, podcasts, YouTube episodes, Loom trials, discovery recordings. Goal-directed watch via Gemini agentic video understanding (frames, audio, or transcrip
Оригинален текст · няма потвърден превод на българскиПърво прочетете инструкцията.
Файлът се зарежда от посочения източник само по ваше действие. Можете да го прочетете тук, да го копирате или да го запазите като .md.
GitHub хранилище- Звезди на репото
- 3559
- Лиценз
- MIT
- Последна проверка на записа
- 22.09.2026 г.
Добавете в своя инструмент.
Командата отваря избора на skills от репозитория. Прочетете файла и условията на автора преди изпълнение.
npx skills add ericosiu/ai-marketing-skillsСтраницата копира текст. Не инсталира и не изпълнява нищо на устройството ви.
Откъс от SKILL.md
Началото на файла на автора, без превод.
Agentic video understanding
Hireable understanding layer. The model takes a goal and decides what to watch, at what speed, and through which modality (frames, audio, transcript), fetching only the moments needed. Vendor claims: up to ~66% lower cost and ~88% fewer tokens vs static fixed-FPS ingest, with higher accuracy.
What this is / is not
Is: goal → watch only what you need → timestamps + quotes + confidence.
Is not: a video editor. Do not cut, overlay, caption-burn, render, schedule, post, email, or write CRM from this skill. Hand cuts to Overlap, FFmpeg, or net-new-video-editor. Approvals stay with the calling lane.
When to use
- Pre-call / sales-call mining: buyer objection, next step, competitive mention
- Shortform scoring: find a 3-second standalone hook and in/out points
- Longform / X research: named-person + contrast moments in podcast or YouTube tape
- Talent review: bar evidence in a Loom or trial recording
- Client audit: every mention of a keyword across a discovery recording
Skip when the job is already a clean transcript and you only need text search.
Inputs
| Field | Required | Notes |
|---|---|---|
source | yes | URL or local media path the runtime can read |
goal | yes | One sentence retrieval goal |
keywords | no | Extra strings to bias retrieval |
max_moments | no | Default 5 |
modality | no | auto (default), frames, audio, or transcript |
Process
- Restate the goal as 1–3 retrieval queries. Done when each query is falsifiable (you would know if a moment matched).
- Call Gemini agentic video understanding (Gemini API or AI Studio) with
source, queries,max_moments, and modality preference. Prefer the agentic path over fixed-FPS full ingest when available. Done when the API returns candidate windows or an explicit empty set. - Normalize moments into the output schema below. Flag paraphrase vs verbatim. Drop fabricated timestamps. Done when every kept moment has
t_start,t_end,modality,quote,why,confidence. - Stop and hand off to the caller. Do not cut, overlay, schedule, publish, email, or CRM-write.
Output schema
Markdown for humans, optional JSON for machines: