Video.from Script

Video From Script

Use this skill when the user wants an avatar video from three concrete inputs: an avatar or image source, a voice id, and approved script text. The backing provider is HeyGen Direct API via the bundled heygen adapter.

Scope

• plan avatar video generation from script text
• start a HeyGen video generation job and return the job id immediately
• poll job status until it is completed, failed, or still rendering
• download the completed MP4 into .generated-videos
• keep HeyGen API keys behind gateway secret injection

Template videos, free-form prompt-to-video, public publishing, social posting, and custom avatar training are outside this skill. Use video-generation for native Sora/Veo prompt videos and heygen for lower-level HeyGen API work. Polling is the completion transport implemented for R55.2; webhook callbacks belong in a provider/gateway adapter if that transport is added later.

Default Workflow

1. Confirm the user has provided an avatar source, voice id, and final script.
2. Run public-facing marketing, sales, training, or onboarding copy through /brand-voice before starting a credit-consuming render.
3. Use plan if the request is vague or needs an explicit risk summary.
4. Use start after the operator grants the exact credit-consuming render. Return the jobId to the user because HeyGen renders asynchronously.
5. Use status --job-id <id> to poll conservatively. Add --download only after the status is complete or when the user asks you to fetch the MP4.
6. Use render --wait only when the user explicitly wants the agent to wait for the final MP4 in the same run.

Do not run parallel render bursts. HeyGen quotas are tight and account/plan dependent.

Command Contract

Run the helper with Node:

node skills/video.from-script/video-from-script.cjs --help

In packaged agent workspaces this skill can be mounted with a hyphenated directory name. If the command above fails with Cannot find module, run:

node skills/video-from-script/video-from-script.cjs --help

Plan without contacting HeyGen:

node skills/video.from-script/video-from-script.cjs plan "Create a product update avatar video"

Start an async render after explicit operator grant:

node skills/video.from-script/video-from-script.cjs start \
  --avatar-id avatar_123 \
  --voice-id voice_123 \
  --script "Approved script text" \
  --title "Product update" \
  --resolution 1080p \
  --aspect-ratio 16:9 \
  --operator-grant

Poll and download the completed MP4:

node skills/video.from-script/video-from-script.cjs status \
  --job-id video_123 \
  --download

Wait for completion in one command only when that behavior is requested:

node skills/video.from-script/video-from-script.cjs render --wait \
  --avatar-id avatar_123 \
  --voice-id voice_123 \
  --script "Approved script text" \
  --operator-grant

Working Rules

• Never print, request, or accept a raw HeyGen API key.
• Keep script text at or below 5000 characters.
• Use exact HeyGen asset ids for --avatar-id and --voice-id. Display names such as a presenter name or voice name are not ids.
• Refresh candidates with node skills/heygen/heygen.cjs request list-avatars --limit <count> and node skills/heygen/heygen.cjs request list-voices --limit <count> before choosing ids. Those summaries are cached so this helper can reject display names and stale ids before contacting HeyGen.
• Use --skip-cache-validation only when the operator supplied a known private HeyGen asset id that is not present in the cached list.
• Provide exactly one avatar source: --avatar-id, --image-url, or --image-asset-id.
• Require --operator-grant for start and render.
• Prefer start + status over a long blocking render.
• Treat pending, waiting, and processing as normal async states.
• Treat failed as terminal and include the provider error when available.
• Download only completed provider URLs, and save MP4 artifacts under .generated-videos.
• In web chat, the completed MP4 must be returned through the helper's artifacts[] output so the gateway can render the browser preview/download route. When asked to post or show an already completed video, run status --job-id <id> --download again instead of writing a remembered local path or hand-built /api/artifact link.
• Do not say web chat cannot embed, display, or deliver the MP4, and do not suggest Finder, drag-and-drop, Discord, or email unless the user explicitly asks for that external channel.
• Public auto-publish or share-link distribution is red tier and requires a separate escalation.

Validation

Run:

python3 skills/skill-creator/scripts/quick_validate.py skills/video.from-script
node skills/video.from-script/video-from-script.cjs --help
node skills/video.from-script/video-from-script.cjs eval-scenarios