produce-video

name	produce-video
model	sonnet
description	Full pipeline: deep research to rendered video
disable-model-invocation	false
argument-hint	<book-slug>

Produce Video

End-to-end video production for Bookie book videos. Takes a book from raw source material to a rendered video in a single conversation, with one human decision: choosing the angle.

This orchestrator delegates each creative phase to its specialist skill and handles the glue between them: voice generation, pipeline execution, content factory, and resume logic.

Working directory: /home/haint/Projects/Bookie/projects/ai-book-video

Input: <book-slug> (e.g., atomic-habits) + author name + book source material (PDF, Drive link, or text)

Resuming

If books/<slug>/ already has files from a previous run, skip completed phases:

• notes.md exists → skip Phase 1, read it and go to Phase 2
• storyboard.md exists → ask if user wants to reuse or regenerate
• chunks.md exists → ask if user wants to reuse or regenerate
• audio/voiceover.wav exists → ask if user wants to reuse or re-voice
• image-prompts.md exists → ask if user wants to reuse or regenerate
• scenes/ has images → produce.sh skips existing images automatically

---

Phase 1: Deep Research + Angle Selection

Invoke /extract-notes <slug>. This handles:

• Knowledge Vault context check (cross-book connections)
• NotebookLM research hub creation (15+ sources across 7 categories)
• Competitive analysis (YouTube + web)
• Structured notes.md output
• 3 angle proposals with template recommendations

Output: books/<slug>/notes.md with chosen angle. The skill ends with angle selection — user picks via AskUserQuestion.

---

Phase 2: Creative Generation

2a. Storyboard

Invoke /create-storyboard <slug>. Uses the chosen angle + recommended narrative template from notes.md.

Output: books/<slug>/storyboard.md with 7-9 scenes, visual concepts, and pacing.

2b. Script

Invoke /write-video <slug>. Produces paired chunk files from storyboard.

Output: books/<slug>/chunks.md (TTS-normalized) + books/<slug>/chunks-display.md (natural Vietnamese for subtitles).

---

Phase 3: Voice Production

3a. Check viXTTS

curl -s --max-time 3 http://127.0.0.1:8020/speakers

If unreachable, tell the user:

viXTTS server is not running. Start it: ./scripts/vixtts-server.sh start Then tell me to continue.

Wait for the user to confirm before proceeding.

3b. Generate Voice

make -C /home/haint/Projects/Bookie/projects/ai-book-video voice BOOK=<slug>

This produces:

• audio/voiceover.wav — the voiceover
• output/section-timing.json — timing authority for all subsequent steps

Report the total duration from section-timing.json.

---

Phase 4: Visual Production

4a. Image Prompts

Invoke /generate-prompts <slug>. Reads section-timing.json + storyboard.md + brand style guide to generate per-scene Gemini prompts.

Output: books/<slug>/image-prompts.md with style-prefixed prompts per scene.

4b. Run Production Pipeline

/home/haint/Projects/Bookie/projects/ai-book-video/scripts/produce.sh <slug> --skip-voice

This runs: images → subtitle → scenes → sync → validate → render.

If it fails, report the error and suggest recovery.

---

Phase 5: Report

After pipeline completes:

• Video file path
• Duration (mm:ss)
• File size
• "Preview your video. Publish when ready."

---

Phase 6: Post-Production

6a. Metadata

Invoke /write-metadata <slug>. Generates YouTube title/description/tags and Facebook caption.

Output: books/<slug>/metadata.md

6b. Catalog Insights

Invoke /catalog-insights <slug>. Catalogs concepts, author profile, and cross-book connections to the Knowledge Vault. This is what makes the system compound — every video makes the next one smarter.

---

Phase 7: Content Factory

After catalog, multiply the video into derivative audio content using NotebookLM MCP. Each artifact is independent — failures don't block others.

Prerequisites: Video rendered, notes.md exists in "Bookie: Library" Master notebook (added in Phase 6b).

1. Get the notebook ID for "Bookie: Library" (cb5c5ce4-4405-44a9-9a94-c7663c896aa9)
2. Create deep-dive podcast:
   • studio_create(artifact_type="audio") with style "deep_dive" focused on this book's angle
   • Poll studio_status every 30s (max 15min timeout)
   • On completion: download_artifact(artifact_type="audio") → books/<slug>/output/podcast-deep-dive.wav
3. Create debate podcast:
   • studio_create(artifact_type="audio") with style "debate" — opposing viewpoints on the book's thesis
   • Poll + download → books/<slug>/output/podcast-debate.wav
4. Create audio briefing:
   • studio_create(artifact_type="audio") with style "briefing" — 5-minute executive summary
   • Poll + download → books/<slug>/output/brief-audio.wav

Report at end:

• Which artifacts succeeded/failed
• File paths and durations
• Total content pieces produced (video + shorts + podcasts + brief)

---

Granular Iteration

If the user wants to redo just one step, point them to individual skills:

• /extract-notes <slug> — redo research + angle
• /create-storyboard <slug> — redo story structure
• /write-video <slug> — redo script
• /generate-prompts <slug> — redo image prompts
• /write-metadata <slug> — redo YouTube/FB copy
• /catalog-insights <slug> — catalog to Knowledge Vault
• make produce BOOK=<slug> — rerun production pipeline