AI Video Processing & Editing with Agent Skills

Step-by-Step Setup

The following setup configures all five skills in your Claude Code environment. The same configuration applies to Cursor and any other MCP-compatible AI assistant.

Step 1: Install FFmpeg

FFmpeg must be installed on your system before the FFmpeg Skill can call it. Verify installation:

ffmpeg -version  # should show version 6.x or higher

Step 2: Configure MCP Skills

Add all five video processing skills to your MCP configuration file at ~/.claude/settings.json:

{
  "mcpServers": {
    "ffmpeg": {
      "command": "npx",
      "args": ["-y", "ffmpeg-mcp-server"]
    },
    "whisper": {
      "command": "npx",
      "args": ["-y", "@openai/whisper-mcp"],
      "env": { "OPENAI_API_KEY": "$OPENAI_API_KEY" }
    },
    "thumbnails": {
      "command": "npx",
      "args": ["-y", "mcp-thumbnail-generator"]
    },
    "youtube": {
      "command": "npx",
      "args": ["-y", "youtube-mcp-server"],
      "env": { "YOUTUBE_CLIENT_ID": "$YT_CLIENT_ID",
                "YOUTUBE_CLIENT_SECRET": "$YT_CLIENT_SECRET" }
    },
    "storage": {
      "command": "npx",
      "args": ["-y", "@aws-sdk/mcp-s3"],
      "env": { "AWS_REGION": "us-east-1",
                "AWS_ACCESS_KEY_ID": "$AWS_ACCESS_KEY_ID",
                "AWS_SECRET_ACCESS_KEY": "$AWS_SECRET_ACCESS_KEY" }
    }
  }
}

Step 3: Authenticate YouTube API

The YouTube API Skill requires OAuth 2.0. Create a project in Google Cloud Console, enable the YouTube Data API v3, and generate OAuth credentials. Run the authentication flow once to obtain a refresh token that the skill will use automatically on subsequent calls.

Step 4: Test Each Skill

• "Transcode sample.mp4 to 720p H.264 and save as output.mp4" — tests FFmpeg Skill
• "Transcribe output.mp4 and save the transcript as transcript.srt" — tests Whisper Skill
• "Generate 3 thumbnail candidates from output.mp4 at 30-second intervals" — tests Thumbnail Skill
• "Upload test.mp4 to my-bucket S3 bucket and return the signed URL" — tests Storage Skill

Automated Workflow: Upload to Publish

Once all five skills are connected, you can automate the complete pipeline with a single agent prompt:

1. Upload — "Upload raw-recording.mp4 to the videos/raw/ prefix in my R2 bucket."
2. Transcode — "Transcode the uploaded file to 1080p H.264, 8 Mbps bitrate, AAC audio at 192 kbps."
3. Transcribe — "Transcribe the processed video and generate an SRT subtitle file and a plain text summary."
4. Generate Thumbnails — "Create 5 thumbnail candidates, overlay the title text from the transcript summary, and return the top-scoring image."
5. Publish — "Upload the processed video to YouTube with the title, description, and tags derived from the transcript. Attach the thumbnail and schedule for 9 AM Eastern tomorrow."

The agent executes each step in sequence, passing output from one skill as input to the next — the transcript text flows into both the thumbnail overlay and the YouTube metadata fields without any manual copy-pasting.

Comparison Table

Use this table to understand which skill handles which stage of the pipeline and what external dependencies each one requires.

Frequently Asked Questions