Scorewright

---

What it does

You give Claude a video. Claude watches it, decides where music starts, stops and hits, chooses a musical identity (genre, key, mode, tempo, instruments, motif) with a reason for each choice, and writes it all in one score.json. One Python script renders that file to real audio:

• Stems (perc, bass, atmos, melodic, sfx), a full mix, a no-percussion mix and composition.mid
• Narration-aware mix: the bed sits 11–12.6 LU under your voice, ducks during words and carves 1.5–4 kHz out of the music
• QC report: loudness, true peak, band energy, hit audibility, stems-sum check, spectrogram

The script has no default style, tempo or key. Every video gets its own music.

Colour is sound

The banner above and the live page use one colour code. Each colour is a part of the spectrum, and so a part of the mix:

Colour	Band	What lives there
🟪 violet #6a3cff	sub, below 60 Hz	sub bass, drops, weight
🟦 blue #3d7bff	bass, 60–250 Hz	chord roots, kick
🟩 teal #19c6b4	250 Hz–1.5 kHz	pads, strings, warmth
⬜ grey #c9cbe0	voice band, 1.5–4 kHz	kept clear for your voice
🟧 amber #ffb02e	upper mids	pulse, bells, motif
🟥 coral #ff4f6d	air, above 5 kHz	hats, sparkle

On the live page, bar height is energy, the grey ribbon lights up while the voice speaks, and a white ring marks a hit point.

Bed or score: the first decision

	Bed	Score
For	narrated explainers, tutorials, Shorts	film, documentary, trailers with speech-free stretches
Melody	none under speech	leitmotif per core idea, developed at story turns
Identity from	timbre and rhythm	melody and harmony
Why	a tune is a second voice: the viewer must choose, and your explanation loses	the music speaks in the gaps

How it works

#	Step	Command / action
0	Check tools	python scripts/scorewright.py doctor
1	Facts	python scripts/scorewright.py analyze video.mp4 work/ --words words.json
2	Watch	Claude reads the contact sheet and the timed words
3	Spot	cue in/out, hit points on words and reveals, where silence is better
4	Direction	genre, key/mode, BPM, palette, leitmotif, dynamic arc — each with a reason
5	Score	Claude writes score.json (format, example)
6	Render	python scripts/scorewright.py render score.json out/v1
7	QC	Claude reads qc.json and spectrogram.png, fixes, re-renders
8	Listen	You listen. Claude cannot hear audio; it fixes what you point at

Craft rules (spotting, leitmotif, emotion mapping, narration-aware mixing) are in references/craft.md.

Install

1. Get the skill into Claude Code

git clone https://github.com/jayaprakash-pixel/scorewright ~/.claude/skills/scorewright
cd ~/.claude/skills/scorewright

2. Python 3.10+ packages

python -m venv .venv
.venv/Scripts/python -m pip install -r requirements.txt     # Windows
# .venv/bin/python -m pip install -r requirements.txt       # macOS / Linux

audio-as-code is optional. Without it, use the gm engine only.

3. External tools

Tool	Get it	Where Scorewright looks
FluidSynth 2.x	releases · brew install fluid-synth · apt install fluidsynth	PATH
GeneralUser GS SoundFont	GeneralUser-GS	soundfonts/GeneralUser-GS.sf2
ffmpeg + ffprobe	ffmpeg.org	PATH

A tool in another place? Copy config.example.json to config.json and set the paths, or set SCOREWRIGHT_FLUIDSYNTH, SCOREWRIGHT_SF2, SCOREWRIGHT_AAC.

4. Check

.venv/Scripts/python scripts/scorewright.py doctor    # prints "ready"
.venv/Scripts/python scripts/test_theory.py           # prints "theory ok"
.venv/Scripts/python scripts/scorewright.py render examples/bed_demo.json out/demo

5. Use it

In Claude Code: /scorewright score this video: my_video.mp4 (add a words.json with word timings for narration-aware mixing).

Known limits

• A General MIDI SoundFont gives a good mock-up, not sample-library realism. Piano, pads, plucks, pizzicato and synths hold up best.
• One tempo per cue. A tempo change is a new cue.
• The aac engine is slow (about 10× the gm engine). Use it for a few sparse tracks.
• Claude cannot listen. The loop is: QC numbers + spectrogram + your ears.

Licence

Scorewright code and docs: MIT © 2026 Jayaprakash, White Noise Academy.

Scorewright does not bundle third-party tools. They keep their own licences: FluidSynth (LGPL-2.1) · GeneralUser GS (GeneralUser GS License v2.0) · audio-as-code (MIT) · ffmpeg (LGPL/GPL) · Three.js (MIT, demo page only).

Music you render with Scorewright is yours. Check the SoundFont licence for your use case.