Rachel Digital Human Production

Rachel Digital Human Production

Core Rule

Use MiniMax for cloned narration and HeyGen for image-driven video. If a MiniMax audio file is available, drive HeyGen with the uploaded audio asset. Do not switch to HeyGen script + voice_id unless the user explicitly asks to use HeyGen's own voice.

Always preserve the production gate:

asset check -> MiniMax narration -> 15-second HeyGen preview -> user approval -> full HeyGen video -> download QA -> archive state

This skill is a workflow guide. It does not grant HeyGen, MiniMax, OpenAI, network, billing, or workspace permissions.

Operating Boundaries

• Read secrets only from MINIMAX_API_KEY and HEYGEN_API_KEY.
• Never print full API keys, Authorization headers, signed temporary URLs, or full request headers.
• Treat MiniMax clone/TTS calls and HeyGen video creation as paid external actions. Request or rely on explicit user approval before the first paid/network action in a run.
• Keep generated assets under inputs/, work/, and outputs/ unless the user names another directory.
• Write or update work/job-state.json before and after every external job stage.
• If a previous state file exists, reuse existing voice_id, asset_id, and video_id when the corresponding source file has not changed.
• If the user asks for a full video, still create a 15-second preview first unless they explicitly waive the preview gate.
• Require the user to confirm that the voice, portrait, script, and final use are authorized before cloning a voice or generating a video.
• Do not provide legal advice. When compliance, disclosure, consent, labor rights, biometric data, or platform policy questions are material, tell the user to verify the applicable rules before publishing.

Required Workflow

1. Ingest

Collect these inputs:

• Script: usually inputs/script.md.
• Portrait: usually inputs/portrait.jpg or inputs/portrait.png.
• Voice source: usually inputs/voice-source.mp3, .m4a, or .wav.
• Optional project name, target platform, aspect ratio, and output naming.

If files live outside the workspace, ask for permission or tell the user what exact file access is needed.

If available, run scripts/preflight_assets.py to check file presence, extensions, sizes, and best-effort media duration before external API calls.

2. Validate Assets

Check before calling any API:

• Voice source format must be MP3, M4A, or WAV.
• Voice source duration must be 10 seconds to 5 minutes.
• Voice source size must be 20 MB or less.
• Portrait must be PNG or JPEG.
• HeyGen normal asset uploads should stay at 32 MB or less.
• The portrait should show a clear front-facing face, visible mouth, and head/shoulders.
• The script should be reviewed for long sentences, brand names, numbers, and acronyms that may need pronunciation testing.

If an input fails validation, stop and propose the smallest fix.

For exact public API facts, read references/api-facts.md. If a user asks for current parameters, pricing, or production code, verify against official MiniMax and HeyGen documentation before taking paid action.

3. Generate MiniMax Narration

• Upload the source voice for cloning.
• Create or reuse a stable voice_id such as brand_person_yyyymmdd.
• Generate the full narration audio with the cloned voice.
• Save the full audio as work/voiceover-full.mp3.
• Save the first 15 seconds as work/preview-15s.mp3.

If no state file exists, use scripts/init_job_state.py or the template in references/checklists.md to create work/job-state.json.

Prefer natural speaking speed for HeyGen mouth tracking. If Chinese lip sync looks strained, remake narration around speed 0.95 to 1.05 before regenerating video.

4. Generate HeyGen Preview

• Upload the portrait as an image asset if no matching image_asset_id exists.
• Upload work/preview-15s.mp3 as an audio asset.
• Submit an Image-to-Video job for a 15-second preview, usually 720p for testing.
• Poll status until completed or failed.
• Download to outputs/preview-15s.mp4.
• Decode-check the downloaded MP4 before marking preview complete.

5. Pause For Approval

After preview generation, stop and ask the user to review:

• voice similarity
• Chinese lip sync
• face shape and mouth distortion
• blinking, head movement, shoulder movement
• framing and platform fit

Do not generate the full video until the user clearly approves the preview.

6. Generate Final

Only after approval:

• Upload work/voiceover-full.mp3 as a HeyGen audio asset if needed.
• Submit the final HeyGen job, usually 1080p.
• Poll status without duplicate job submission.
• Download to outputs/final-1080p.mp4 or the user-specified output name.
• Use retry/resume for downloads when available.
• Decode-check the entire MP4, not just the first few seconds.
• Update work/job-state.json.

Batch Mode

For multiple scripts:

• Create one state record per video.
• Generate previews only by default.
• Place previews under outputs/previews/.
• Produce a review sheet listing preview path, duration, video_id, and review notes.
• Do not generate any full videos until the user approves specific previews.

Failure Policy

• If an API call returns a clear failed status, record the failure reason and ask before retrying paid generation.
• If polling times out, do not assume generation failed. Resume polling by video_id.
• If download fails or the MP4 is corrupt, retry download before regenerating the video.
• If a signed URL expires, refetch job status or asset metadata instead of creating a new paid job.

Public Distribution Policy

• Treat this skill as shareable workflow instructions, not as a hosted service.
• Do not bundle .env files, API keys, private portraits, voice samples, generated videos, job-state files from real clients, or signed download URLs.
• If publishing the skill publicly, include only SKILL.md, agents/openai.yaml, references/, and scripts/.
• Make the user bring their own MiniMax and HeyGen accounts, keys, billing, and rights to the source materials.
• For public use, prefer explicit invocation with $rachel-digital-human-production so paid workflows do not trigger accidentally.
• If publishing through a repository, marketplace, or team distribution channel, choose a package-level license outside the skill instructions.

References

Read these only when needed:

• references/checklists.md for the state template, reusable prompts, and QA checklist.
• references/api-facts.md for the currently documented MiniMax and HeyGen facts this workflow relies on.
• references/public-safety.md for consent, disclosure, sharing, and public-release guidance.

If exact API parameters, pricing, availability, or provider behavior matters, verify against official MiniMax and HeyGen documentation before coding or running paid requests.