write-video

name	write-video
model	sonnet
description	Write book video script as paired TTS-ready chunks
disable-model-invocation	false
argument-hint	<book-slug>

Write Video Script

Write a Bookie book video script directly as numbered chunks, outputting paired files for display (subtitles) and TTS (voice generation). This replaces the old write-script + split-script + estimate flow.

Context

• When to use: After /create-storyboard produces storyboard.md. Before make voice.
• Input: storyboard.md (story direction) + notes.md (insights) + tts-config.json (normalization rules)
• Output: chunks-display.md (readable script + subtitle source) + chunks.md (TTS-normalized)

Pipeline Position

/create-storyboard → storyboard.md
        ↓
/write-video → chunks-display.md + chunks.md  ← YOU ARE HERE
        ↓
make voice → audio + section-timing.json
        ↓
/generate-prompts → image-prompts.md

Steps

1. Validate: Check that $ARGUMENTS is provided. If missing, ask Hai for the book slug. Set SLUG=$ARGUMENTS. Check that projects/ai-book-video/books/$SLUG/ exists.

2. Read input files:

   • Read projects/ai-book-video/books/$SLUG/storyboard.md — scene plan, story arc, pacing
   • Read projects/ai-book-video/books/$SLUG/notes.md — book title, author, insights, angle, stories
   • Read projects/ai-book-video/tts-config.json — chunk params and en_to_vn dictionary
   • Read projects/ai-book-video/brand/style-guide.md — tone reference

   If storyboard.md is missing, warn and ask if Hai wants to proceed without it (using notes.md only) or run /create-storyboard first.

3. Detect mode:

   • If chunks-display.md exists with actual narration content → revision mode
   • If missing or empty → new draft mode

   Read references/writing-guide.md — chunk-aware writing guidelines, scene structure, tone/style, revision mode procedure. Read for both new draft and revision modes.

4. Write display chunks (new draft): Follow the storyboard's scene plan. Write narration as numbered chunks [NNN] "text". Each chunk is one natural speech unit — a complete thought that sounds right when spoken aloud.

5. Generate TTS chunks: For each display chunk, create its TTS-normalized version.

   Read references/format-rules.md — chunk format specs, TTS normalization rules (numbers, EN terms, avoid patterns), chunk size validation, file headers. Read before writing output files.

6. Validate TTS chunks: Check each TTS chunk's character count (Python len()). Sweet spot: 75-250 chars. Flag outliers.

7. Write both files to projects/ai-book-video/books/$SLUG/

8. Print summary: Total chunks, total words, per-scene breakdown, normalizations applied, warnings, next step: make voice BOOK=$SLUG

Output Files

chunks-display.md — Human-readable script AND subtitle source. Contains scene headings (## HOOK, ## CONTEXT...), voice config, **[SHORT]** markers.

chunks.md — TTS-normalized version. Same structure, headings are ## Scene 01, normalized text. This is what viXTTS speaks.

Read references/format-rules.md for exact file header templates, scene marker syntax, and numbering conventions.

Important

• Both files have identical chunk numbering — [001] in display maps to [001] in TTS
• Scene markers and pace values identical in both files
• The display file IS the readable script — there's no separate script.md
• Vietnamese text: count chars with len() (Python), not byte count
• After writing, next step: make voice BOOK=$SLUG