Analyze Footage — AI Video Understanding

You are helping the user understand their existing video files using TwelveLabs MCP. This gives your agent "eyes" to see what's in video footage.

Load the remotion-production skill for the video-analysis rule.

Workflow

1. Identify Video Files

Ask the user which video files to analyze, or look for them:

# Check common locations
ls public/footage/ public/raw/ public/video/ public/ 2>/dev/null | grep -E '\.(mp4|mov|avi|mkv|webm)$'

If no files found, ask the user where their footage is located.

2. Upload to TwelveLabs

For each video file, upload it as an asset (no index needed for analysis):

create-asset:
  file: [absolute path, e.g. /path/to/project/public/footage/source-video.mp4]

Keep the returned asset ID. If the next call says the asset isn't ready yet, wait a few seconds and retry.

3. Analyze Content

Scene Breakdown — a timestamped scene map, as the video-analysis rule shows:

async-analyse-video:
  assetId: [asset ID]
  modelName: "pegasus1.5"
  analysisMode: "time_based_metadata"
  segmentDefinitions: [one "shots" definition with fields: description, shot_type, people, on_screen_text, usable]

Poll get-analyse-task every 5-10 seconds until it's ready, then parse result.data (a JSON string keyed by the segment definition id). Videos up to 2 hours work.

For a direct question ("which shots show the product?", "pick clips for a 30s promo"), use sync-analyse-video (assetId, modelName: "pegasus1.5", prompt, and a jsonSchema so you get start/end times back as numbers). Sync analysis is capped at 1 hour of video.

Always pass modelName: "pegasus1.5": the server's default model has been removed from the API.

Present as:

📹 Video Analysis: [filename]

Scenes:
  0:00 - 0:15  │ Logo animation, dark background
  0:15 - 0:45  │ Speaker at podium, introducing topic, conference setting
  0:45 - 1:20  │ Product demo on laptop screen, close-up
  1:20 - 1:45  │ Audience reaction, medium shot, clapping
  1:45 - 2:00  │ Call to action slide, company branding

Key Elements (from the segment fields):
  - People: 1 speaker, audience (~30 people)
  - Shot types: wide, medium, close-up, screen capture
  - Text on screen: company logo, product name
  - Setting: conference room, professional lighting

When the results are saved (e.g. to analysis/[filename].json), clean up with delete-asset (assetId). Assets count against the account's storage.

4. Targeted Search (if user has specific needs)

For one file, ask sync-analyse-video directly (optionally clipped with startTime/endTime). To find moments across a library of videos, index it once and search:

start-video-indexing-task:
  folderFilePath: [absolute path to the footage folder]   (or videoFilePath for one file)
→ poll get-indexed-asset until ready
search:
  query: "the part where the product is demonstrated"

search returns ranked segments with start/end times and thumbnails; it doesn't describe them. Analyse a hit if you need to know more.

5. Recommend Clips

Based on the analysis, recommend the best clips for the user's video:

🎬 Recommended Clips for a 30-second promo:

1. Opening (0:00-0:05)
   → Use 0:45-0:50 of source — product demo close-up, high energy
   
2. Main content (0:05-0:20)  
   → Use 0:50-1:05 of source — full demo walkthrough
   
3. Social proof (0:20-0:25)
   → Use 1:20-1:25 of source — audience clapping
   
4. CTA (0:25-0:30)
   → Use 1:45-1:50 of source — branded slide

6. Show Usage in Remotion

For each recommended clip, show the Remotion code. A source range becomes trimBefore (the start timestamp) plus durationInFrames (the clip length) on <Video> from @remotion/media. Cut the clips together with TransitionSeries or a Track as the official remotion-markup skill shows (video-editing.md); skills/remotion-production/rules/video-analysis.md has a full example.

import { Video } from "@remotion/media";
import { staticFile, useVideoConfig } from "remotion";

const { fps } = useVideoConfig();

// Opening (0:00-0:05) ← source 0:45-0:50
<Video
  name="Opening"
  src={staticFile("footage/source-video.mp4")}
  trimBefore={45 * fps}
  durationInFrames={5 * fps}
  objectFit="cover"
  style={{ width: "100%", height: "100%" }}
/>

Round fractional TwelveLabs timestamps to whole frames: Math.round(45.6 * fps).

Tips

• Ensure video files are in public/ (ideally public/footage/) so Remotion can access them via staticFile() with the path relative to public/; pass the absolute path of the same file to create-asset
• Limits: sync analysis up to 1 hour, async up to 2 hours, direct uploads up to 4 GB
• Index only for search; an index is reusable across many searches, an asset is per analysis (delete it afterwards)
• For anything else, see the official docs: https://docs.twelvelabs.io (and llms.txt)
• Combine analyzed footage with stock clips from Pexels to fill gaps