Document to Narration

Convert written documents into narrated video scripts with precise word-level timing.

Core Principle

The agent interprets; the document guides. Rather than rigid template-based splits, this skill uses agent judgment to find where the content naturally breathes, argues, and transitions. The document's argument flow determines scene breaks, not a predetermined structure.

When to Use This Skill

Use this skill when:

• Converting a blog post or essay to video narration

• Preparing content for TTS audio generation

• Breaking long-form content into digestible scenes

• Creating word-level synchronized captions for video

Do NOT use this skill when:

• The content is already in scene/script format

• You need real-time voice synthesis (this is batch processing)

• Working with dialogue or multi-speaker content (single voice only)

Prerequisites

• Deno installed (https://deno.land/)

• Python 3.12 with venv support

• ffmpeg for audio conversion

• whisper-cpp (installed via @remotion/install-whisper-cpp)

• TTS model at tts/model/ (not in git due to size - see Model Setup below)

Complete Pipeline

There are two approaches: per-scene (legacy) and full narration (recommended).

Full Narration Pipeline (Recommended)

Generates a single audio file for consistent volume and pacing:

Document (.md) ↓ [agent interprets scene breaks] Scene .txt files (01-scene-name.txt, 02-scene-name.txt, ...) ↓ [TTS via narrate-full.py - SINGLE PASS] full-narration.wav (one consistent audio file) ↓ [Whisper via transcribe-full.py] full-narration.json + full-narration.vtt (word-level timing) ↓ [extract-scene-boundaries.py] Scene timing boundaries for video composition

Per-Scene Pipeline (Legacy)

Generates separate audio per scene - can cause volume inconsistencies:

Scene .txt files ↓ [TTS via narrate-scenes.py - MULTIPLE PASSES] Scene .wav files (volume may vary between scenes) ↓ [concatenate] Combined audio (may have clipping at boundaries)

Warning: Per-scene TTS generates audio with different volume levels and pacing. When concatenated, this causes audible jumps and clipping. Use the full narration pipeline instead.

Quick Start

Full Narration Pipeline (Recommended)

cd .claude/skills/document-to-narration source tts/.venv/bin/activate

1. Split document into scenes (manual or scripted)

deno run --allow-read --allow-write scripts/split-to-scenes.ts input.md --output ./output/

2. Generate single audio file

python scripts/narrate-full.py ./output/scenes/

3. Transcribe with word-level timestamps

python scripts/transcribe-full.py ./output/full-narration.wav

4. Extract scene boundaries for video timing

python scripts/extract-scene-boundaries.py ./output/scenes/ ./output/full-narration.json --typescript

Legacy Per-Scene Pipeline

1. Split document into scenes

deno run --allow-read --allow-write scripts/split-to-scenes.ts input.md --output ./output/

2. Generate audio per scene (may have volume inconsistencies)

source tts/.venv/bin/activate python scripts/narrate-scenes.py ./output/scenes/

3. Transcribe (DEPRECATED: transcribe-scenes.ts requires whisper-cpp)

Use transcribe-full.py instead after concatenating audio

Instructions

Step 1: Setup (First Time Only)

Create Python Virtual Environment

cd .claude/skills/document-to-narration/tts python3.12 -m venv .venv source .venv/bin/activate pip install -r requirements.txt

TTS Model Setup

The fine-tuned voice model (~7.8GB) is not included in git due to size. Place your Qwen3-TTS model files in tts/model/ :

tts/model/ ├── config.json ├── generation_config.json ├── model.safetensors # Main model weights ├── tokenizer_config.json ├── vocab.json ├── merges.txt └── speech_tokenizer/ └── ...

Install Whisper (if not already installed)

The @remotion/install-whisper-cpp package handles this:

import { installWhisperCpp, downloadWhisperModel } from '@remotion/install-whisper-cpp';

await installWhisperCpp({ to: './whisper-cpp', version: '1.5.5' }); await downloadWhisperModel({ model: 'medium', folder: './whisper-cpp' });

Step 2: Prepare Your Document

The skill works best with:

• Markdown documents with clear heading structure (H1, H2)

• Well-structured arguments with distinct sections

• Content that reads naturally aloud

Step 3: Run the Pipeline

deno run -A scripts/full-pipeline.ts /path/to/essay.md --output ./output/essay-name/

Step 4: Review Output

output/essay-name/ ├── scenes/ │ ├── 01-opening-hook.txt # Scene script │ ├── 01-opening-hook.wav # Generated audio │ ├── 01-opening-hook.vtt # Word-level captions │ ├── 02-core-argument.txt │ ├── 02-core-argument.wav │ ├── 02-core-argument.vtt │ └── ... └── manifest.json # Complete timing data

Scene Boundary Heuristics

The agent identifies scene breaks using these heuristics:

Strong Boundaries (Almost Always Break)

• H2 heading changes

• "Here's the thing" / "The point is" pivot statements

• Major metaphor introduction

• Explicit enumeration ("First...", "Second...")

• Significant perspective shifts

Moderate Boundaries (Consider Breaking)

• Long paragraph after short ones (or vice versa)

• Example-to-principle transitions

• "But" / "However" / "Meanwhile" at paragraph start

• Question-then-answer patterns

Weak Boundaries (Usually Keep Together)

• Paragraph-to-paragraph within same example

• Sequential evidence for same point

• Build-up to a punchline/reveal

Scene Length Guidance

• Target: 100-300 words per scene (30-90 seconds of audio)

• Minimum: 50 words (avoid micro-scenes)

• Maximum: 500 words (avoid cognitive overload)

Anti-Patterns

The Paragraph Slicer

Pattern: Breaking at every paragraph or heading mechanically. Problem: Ignores argument flow. Scenes feel choppy and disconnected. Fix: Look for rhetorical units, not structural units. Multiple paragraphs often form one scene.

The Wall of Text

Pattern: Keeping entire sections as single scenes. Problem: Creates TTS audio that's too long. Loses natural breathing room. Fix: Target 100-300 words. Find the natural pause point within sections.

The Verbatim Transcriber

Pattern: Copying written text exactly without spoken adaptation. Problem: Written conventions don't work when spoken. Parentheticals, complex punctuation, and nested clauses confuse TTS and listeners. Fix: Apply adaptation rules. Read it aloud mentally.

The Over-Adapter

Pattern: Rewriting content so heavily it loses the author's voice. Problem: The result doesn't sound like the original author. Fix: Preserve voice, adjust mechanics. If the author uses rhetorical questions, keep them.

Available Scripts

scripts/split-to-scenes.ts

Parse a markdown document and output scene text files.

deno run --allow-read --allow-write scripts/split-to-scenes.ts input.md --output ./output/ deno run --allow-read --allow-write scripts/split-to-scenes.ts input.md --output ./output/ --adapt deno run --allow-read scripts/split-to-scenes.ts input.md --dry-run

Options:

• --output

• Directory for scene files (created if doesn't exist)

• --adapt

• Apply spoken adaptation rules

• --dry-run

• Preview scene breaks without writing files

Output: Numbered .txt files and initial manifest.json

scripts/narrate-full.py (Recommended)

Generate a single TTS audio file from all scene files. Produces consistent volume and pacing.

python scripts/narrate-full.py ./output/scenes/ python scripts/narrate-full.py ./output/scenes/ --force python scripts/narrate-full.py ./output/scenes/ --speaker jwynia python scripts/narrate-full.py ./output/scenes/ --output ./custom/path/audio.wav

Options:

• --force

• Regenerate even if output exists

• --speaker

• Speaker name (default: jwynia)

• --output

• Custom output path (default: ../full-narration.wav )

Output: Single full-narration.wav in parent directory of scenes

scripts/narrate-scenes.py (Legacy)

Generate TTS audio for each scene file separately. Not recommended - can cause volume inconsistencies when concatenated.

python scripts/narrate-scenes.py ./output/scenes/ python scripts/narrate-scenes.py ./output/scenes/ --force python scripts/narrate-scenes.py ./output/scenes/ --speaker jwynia

Options:

• --force

• Regenerate even if output exists

• --speaker

• Speaker name (default: jwynia)

Output: .wav files alongside each .txt file

scripts/transcribe-full.py (Recommended)

Transcribe audio with word-level timestamps using Python's openai-whisper.

python scripts/transcribe-full.py ./output/full-narration.wav python scripts/transcribe-full.py ./output/full-narration.wav --model large-v3 python scripts/transcribe-full.py ./output/full-narration.wav --output-dir ./captions/

Options:

• --model

• Whisper model: tiny, base, small, medium, large, large-v2, large-v3 (default: medium)

• --output-dir

• Output directory (default: same as audio file)

Output:

• .vtt file with word-level timestamps

• .json file with captions array for Remotion

Dependencies: Requires openai-whisper in Python environment:

pip install openai-whisper

scripts/extract-scene-boundaries.py

Extract scene timing boundaries from transcript by matching scene opening phrases.

Human-readable table

python scripts/extract-scene-boundaries.py ./output/scenes/ ./output/full-narration.json

JSON output

python scripts/extract-scene-boundaries.py ./output/scenes/ ./output/full-narration.json --json

TypeScript for Video.tsx

python scripts/extract-scene-boundaries.py ./output/scenes/ ./output/full-narration.json --typescript

Options:

• --json

• Output as JSON array

• --typescript

• Output as TypeScript code for Video.tsx scenes array

Output: Scene numbers, slugs, start times, and durations

scripts/transcribe-scenes.ts (Deprecated)

Deprecated: Requires whisper-cpp binary which may not be installed. Use transcribe-full.py instead.

Transcribe per-scene audio files using whisper-cpp.

deno run --allow-read --allow-write --allow-run scripts/transcribe-scenes.ts ./output/scenes/

Output: .vtt files with word-level timestamps

scripts/full-pipeline.ts

Orchestrate the complete pipeline.

deno run -A scripts/full-pipeline.ts input.md --output ./output/project-name/

Options:

• --output

• Output directory (required)

• --adapt

• Apply spoken adaptation

• --skip-tts

• Skip audio generation (text only)

• --skip-transcribe

• Skip Whisper transcription

Output Format

manifest.json

{ "source": "appliance-vs-trade-tool-draft.md", "created_at": "2024-01-15T10:30:00Z", "total_scenes": 9, "total_duration_seconds": 420, "scenes": [ { "number": 1, "slug": "popcorn-opening", "word_count": 185, "audio_duration_seconds": 55.2, "files": { "text": "scenes/01-popcorn-opening.txt", "audio": "scenes/01-popcorn-opening.wav", "captions": "scenes/01-popcorn-opening.vtt" }, "captions": [ { "text": "Two", "startMs": 0, "endMs": 180, "confidence": 0.98 }, { "text": "people", "startMs": 180, "endMs": 450, "confidence": 0.97 } ] } ] }

VTT Format

WEBVTT

00:00.000 --> 00:00.180 Two

00:00.180 --> 00:00.450 people

00:00.450 --> 00:00.720 walk

00:00.720 --> 00:01.100 into

Spoken Adaptation

When --adapt is enabled, the skill transforms written conventions to spoken equivalents:

Written Spoken

Parenthetical asides Em-dash or separate sentence

"e.g." "for example"

"i.e." "that is"

Long nested clauses Split into multiple sentences

Semicolons Periods

emphasis

Context-appropriate stress

Preserve:

• Author's voice and tone

• Rhetorical questions

• Deliberate repetition

• Key phrases and memorable formulations

Integration

With remotion-designer

• Pass manifest scene list to remotion-designer

• Each scene becomes a visual design unit

• Word-level timing drives text animation

With Remotion Compositions

import { Audio, useCurrentFrame, Sequence } from 'remotion'; import manifest from './output/manifest.json';

// Use scene durations for Sequence timing {manifest.scenes.map((scene, i) => ( <Sequence from={accumulatedFrames} durationInFrames={scene.audio_duration_seconds * fps}

<Audio src={staticFile(scene.files.audio)} />

<CaptionRenderer captions={scene.captions} />

</Sequence> ))}

Technical Notes

WAV Format Conversion

Whisper requires 16kHz mono WAV. The pipeline handles conversion automatically:

ffmpeg -i input.wav -ar 16000 -ac 1 output_16khz.wav

TTS Model

The fine-tuned voice model (~7.8GB) is bundled at tts/model/ . Uses Qwen3-TTS with custom speaker embedding.

Performance

• TTS: ~5-30 seconds per sentence (Apple Silicon MPS or NVIDIA CUDA)

• Whisper: ~0.5-2x realtime depending on model size

• Full essay (~2000 words): ~10-20 minutes total processing

What This Skill Does NOT Do

• Generate video visuals (use remotion-designer)

• Real-time voice synthesis

• Multi-speaker dialogue

• Edit or improve the content's argument

• Make editorial changes beyond mechanical spoken adaptation