demo-video

name: demo-video description: Generate a narrated demo video from browser screenshots and TTS audio. Captures scenes via agent-browser, generates voice narration via Inworld TTS API, and stitches into an MP4 with ffmpeg. #

Generate a narrated demo video from browser screenshots and TTS audio. The output is an MP4 with still-image scenes synced to voice narration.

Prerequisites #

• agent-browser installed
• ffmpeg installed
• Inworld API key in api/.dev.vars (INWORLD_API_KEY)

Process #

Step 1: Take Screenshots

Use agent-browser to navigate the app and capture screenshots. Save all files to tmp/video/ in the project root.

Example:

scene1_overview.png — Navigate to main page, full page screenshot
scene2_feature.png — Interact with a feature, capture result
scene3_detail.png — Click into detail view, capture

Step 2: Build the Video (use a subagent)

IMPORTANT: Once all screenshots are captured and narration text is written, delegate the TTS generation, ffmpeg stitching, and R2 upload to a background subagent using the Task tool. This keeps the main conversation responsive and avoids filling the context window with ffmpeg output.

The subagent prompt should include:

• The list of screenshot filenames in tmp/video/
• The narration text for each scene
• Instructions to generate TTS audio, create video segments, concatenate, upload to R2, and open the final video

Below are the details the subagent needs:

TTS Audio Generation

Call the Inworld TTS sync API to generate MP3 audio files for each scene.

API endpoint: POST https://api.inworld.ai/tts/v1/voice (non-streaming, returns complete audio in one response)

Request:

INWORLD_API_KEY=$(grep '^INWORLD_API_KEY=' api/.dev.vars | cut -d'=' -f2-)
curl -s -X POST "https://api.inworld.ai/tts/v1/voice" \
  -H "Authorization: Basic ${INWORLD_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Your narration text here.",
    "voiceId": "layercode_production__el_5",
    "modelId": "inworld-tts-1.5-max",
    "audioConfig": {
      "audioEncoding": "MP3",
      "sampleRateHertz": 22050
    }
  }' | jq -r '.audioContent' | base64 --decode > tmp/video/scene1_audio.mp3

CRITICAL — Inworld rate limit: The Inworld API rejects concurrent requests with a misleading SESSION_TOKEN_INVALID error (gRPC code 16). Generate TTS files sequentially, not in parallel. Chain them with && in a single bash command.

Text limit: 2,000 characters per request. Split longer narration across multiple calls.

Validation: Always check file sizes after generation. Valid MP3 files are 50KB+. Files of 3 bytes or less indicate the API returned an error — check the raw response with curl ... | head -c 500.

ffmpeg Video Segments

For each scene, create a video that displays the screenshot for the duration of its audio:

ffmpeg -loop 1 -i scene1_overview.png -i scene1_audio.mp3 \
  -c:v libx264 -tune stillimage -c:a aac -ar 22050 -b:a 128k \
  -pix_fmt yuv420p -shortest -y scene1.mp4

Concatenate into Final Video

printf "file 'scene1.mp4'\nfile 'scene2.mp4'\nfile 'scene3.mp4'\n" > concat.txt
ffmpeg -f concat -safe 0 -i concat.txt -c copy -y demo.mp4

Final output: tmp/video/demo.mp4

Run open tmp/video/demo.mp4 to show the final video to the user.

Upload Video to R2

Upload the final video to the toyo-dev-demo-videos R2 bucket using wrangler. The bucket has public access enabled so videos can be shared via URL.

FILENAME="$(git branch --show-current).mp4"

cd api && npx wrangler r2 object put "toyo-dev-demo-videos/${FILENAME}" \
  --file ../tmp/video/demo.mp4 \
  --content-type "video/mp4" \
  --remote

Other R2 commands:

cd api && npx wrangler r2 object list toyo-dev-demo-videos --remote    # List videos
cd api && npx wrangler r2 object delete "toyo-dev-demo-videos/${FILENAME}" --remote  # Delete

Public URL format: https://pub-<hash>.r2.dev/<filename> (if public access is configured on the bucket).