Reel Maker

Reel Maker

Create production-ready YouTube Shorts / Instagram Reels with animated motion graphics, timed captions, and H.264 MP4 export.

What This Skill Produces

Given a topic, outline, or script → this skill outputs:

1. A 6-scene narration script with timestamped caption chunks
2. An animated React preview (optional, for timing/pacing review)
3. A 1080×1920 @30fps H.264 MP4 with animated visuals
4. A narrated MP4 with TTS audio muxed in (Dia, Higgs Audio, or Flite)

Dependencies

Core (always required):

• Python 3 with Pillow (pip install pillow --break-system-packages if needed)
• FFmpeg (system-installed, verify with which ffmpeg)
• DejaVu fonts (standard on Ubuntu; on macOS/Windows the renderer falls back to system fonts automatically)

TTS — install one (in order of quality):

• Dia (recommended): pip install git+https://github.com/nari-labs/dia.git + CUDA GPU (~4.4GB)
• Higgs Audio API: pip install requests + set DEEPINFRA_API_KEY env var
• Flite (fallback): libflite system package (pre-installed on Ubuntu)

Preview (optional):

• React artifact with Tone.js (available in Claude artifacts)

Workflow

Phase 1: Script

Read references/script_rules.md before writing the script.

Input: A topic, video outline, or brief description. Output: A markdown file containing:

- Scene breakdown (6 scenes with timestamps, labels, colors)
- Full narration (~155-170 words for 55-60 seconds)
- Caption chunks array: [startTime, endTime, "3-5 words"] per chunk
- Cross-link target (which video the cliffhanger teases)

The 6-scene structure is always:

Phase 2: Animated Preview (Optional)

If the user wants to preview before rendering, create a React JSX artifact:

• Each scene is a component receiving progress (0→1)
• Captions displayed as 3-5 word chunks in a pill overlay at the bottom
• Segmented progress bar at top, color-coded per scene
• Tone.js audio: ambient drone (PolySynth FM, C2+G2, -26dB through LowPass 350Hz + Reverb), scene transition hits (MembraneSynth), caption ticks (MetalSynth, -32dB)
• Controls: play/pause, restart, mute, scene jump buttons

Important: speechSynthesis is blocked in artifact iframes. Use only Tone.js for audio. Voiceover is recorded separately.

Phase 3: Render to MP4

This is the production pipeline. Read references/export_specs.md for the locked FFmpeg command and codec specs.

Step 1: Import shared helpers:

# Copy core/drawing.py and core/scene_base.py to working directory
# These provide: draw_rounded_rect, centered_text, left_text, right_text,
#                rgba, draw_progress_bar, draw_caption, render_frame shell

Step 2: Write scene renderers. Each scene is a function:

def draw_scene_N(draw, img, progress):
    # progress is 0.0 → 1.0 across the scene's duration
    # Use draw (ImageDraw) and img (PIL Image) to render
    # All animation derived from progress via easing math

For animation patterns (glitch effects, filling containers, charts, diagrams, crack effects), read references/scene_cookbook.md.

For a complete working example, read examples/video5_context_erosion.py.

Step 3: Render and pipe to FFmpeg:

from core.renderer import render_to_mp4
render_to_mp4(scenes, captions, scene_renderers, output_path)

Step 4: Validate output:

ffprobe -v quiet -print_format json -show_streams output.mp4

Verify: codec=h264, 1080×1920, 30fps, yuv420p, correct frame count.

Design System (Locked)

Common Animation Patterns

Quick reference (full details in references/scene_cookbook.md):

fade_in     = min(1, progress * 3)                        # Fast fade
delayed     = max(0, (progress - 0.5) * 2)                # Appears at 50%
slide_up    = (1 - opacity) * 20                           # Y offset
pulse       = 0.5 + math.sin(progress * 20) * 0.5         # Oscillation
cliff_drop  = max(0, min(1, (progress - 0.35) * 2.5))     # Delayed sharp drop

Phase 4: Narration (TTS)

Read references/tts_setup.md for installation of TTS backends.

The skill auto-detects the best available TTS engine:

Generate narration and mux into MP4:

from core.tts import generate_narration, mux_audio_video

# Each scene dict needs 'id', 'start', 'end', 'text' keys
narration_wav = generate_narration(scenes, output_dir="/output",
                                    voice_seed=42)  # seed = consistent voice
mux_audio_video("silent_video.mp4", narration_wav, "final_with_audio.mp4")

The pipeline generates speech per scene, time-stretches each to fit its scene duration, aligns them to exact timestamps via FFmpeg adelay, and muxes the result into the MP4 as AAC audio.

Dia speaker tags: For single-narrator Shorts, all text is wrapped in [S1]. Non-verbal tags like (laughs), (sighs), (clears throat) are supported but should be used sparingly.

Voice consistency: Dia generates a random voice per run. Use a fixed voice_seed to keep the same voice across all 6 scenes. Change the seed for a different voice.

Known Limitations

Post-Production (If Using Flite or No TTS)

If the narration was generated with Dia or Higgs Audio, the MP4 is ready for upload. If using Flite or no TTS:

1. Import MP4 into editor (Premiere / DaVinci / CapCut)
2. Record voiceover at ~160 WPM matching caption timing
3. Layer voiceover + optional music bed (-20dB under voice)
4. Export H.264 1080×1920 30fps matching source
5. Upload: YouTube Shorts (pin comment with full video link), Instagram Reels (same file)