Scorewright

A copilot for composers. Private project.

Brian composes, and has 1,607 of his own scores exported from Sibelius to MusicXML. Scorewright lets Claude Code compose with him the way it codes with a programmer. It reads his whole score, pulls examples from his own archive, writes new bars in ABC, splices them into his MusicXML, and shows them in green on a phone-friendly page. He listens, then accepts or rejects with a spoken reason. Every verdict is logged, and those verdicts are the evaluation set the harness is tuned against.

There is no trained music model in the loop. Training was tried first (NotaGen, Text2Score, a $50 GPU day). Its continuations scored 0 or 1 out of 5; the one exception was a four-bar left-hand fill that got a 3. Claude reading the whole piece and writing the bars itself scored 3 and 4 the same day. docs/history.md tells that story.

How it works

• lab/app/server.py runs under launchd and serves the page on port 8790, on the Mac and on the phone through Tailscale.
• Brian talks, types, plays (🎹 with a MIDI keyboard or the mic), hums (🎵) or uploads a recording (📎). Everything he says lands in _bridge/queue/inbox.jsonl.
• A Claude Code session waits on that inbox, composes by the procedure in lab/COMPOSING.md, and registers the result as a suggestion.
• The page draws the suggestion with Verovio, the new bars green, and plays it: MuseScore audio within seconds, NotePerformer through Dorico 4 a minute or so later.
• Accept makes it the working score; Reject keeps the old one. Both go to _bridge/verdicts.jsonl with his words.

Quick start

For Brian:

1. Open http://localhost:8790 on the Mac, or https://macbook-pro.tail2783bf.ts.net on the phone with Tailscale on.
2. In Claude Code, in ~/repos/scorewright, ask it to start the Scorewright loop. It follows the checklist in CLAUDE.md and starts two background helpers: the inbox waiter and the NotePerformer renderer.
3. Pick a piece in the header menu, or "New piece" for an empty score.
4. Say what you want, or press ▶ continue. Accept or Reject each suggestion, then say why.
5. "open in Dorico" opens what you are looking at in Dorico 4, green notes kept, with NotePerformer.

For a Claude Code session: read CLAUDE.md, then lab/COMPOSING.md.

Requirements

• macOS, with the conda envs whisperer (nearly everything) and text2score (human_apply.py, dump_abc.py), plus the anaconda base python for hum.py. Never a bare python.
• Dorico 4 with NotePerformer 5 (its default playback template), MuseScore 4, Tailscale.
• The data folder ~/Documents/Score Exports: the archive, the retrieval index, the session folders, the bridge files. It is not in the repo.

Layout

path	what
CLAUDE.md	operating manual for Claude Code sessions
lab/COMPOSING.md	the composing procedure and the rules learned from his verdicts
lab/app/	the server, the page, rendering, NotePerformer, input transcription, the MuseScore panel
lab/*.py	composing tools, plus scripts from the training phase
lab/band/	the concert band orchestration of Bright Valley
lab/songs/	build scripts for whole new pieces: where_i_left_it.py (S1, too pop), where_i_left_it_s2.py (S2, odd), where_i_left_it_s3.py (S3, in his song-study grammar), English horn and piano with words
lab/style_profile.py	measures a composer's harmonic and melodic habits from their scores and flags where a draft falls outside them
lab/rh_profile.py	what the accompaniment's right hand does against the melody: doubling, thirds and sixths, fills in the melody's rests
lab/music_map.py	the Phrase Atlas: every phrase of a set of scores mapped by melody, rhythm and harmony (UMAP), with named regions and nearest neighbours; lab/app/phrase_atlas.html is the viewer
lab/engrave_pages.py	engraves a score to page images, so the whole piece is re-read as pages before every continue or revision
lab/lint_line.py	checks a melody for wrong-way accidentals, unrecovered or augmented leaps, phrases too long to breathe, range
lab/harmony_index.py, lab/find_progression.py	index the harmony of all 1,655 scores; find a progression in any key and show the melody he wrote over it
lab/reduce_harmony.py	prints a score as a lead-sheet reduction, one line per bar (chords, bass, melody, accompaniment rhythm), to read many pieces fast
lab/NotaGen/	git submodule; its ABC converters are used by the pipeline
*.py at the root	xmledit.py, diffview.py, retrieve.py, ranges.py, analyze.py, and the August prototypes
deploy/	the launchd plist
docs/	the documentation below

Docs

• docs/architecture.md: components, data flow, state.json, HTTP routes
• docs/surface.md: the page and the MuseScore panel
• docs/playback.md: MuseScore and NotePerformer audio
• docs/input.md: talk, take, hum, upload
• docs/corpus.md: the archive, retrieval, the export history
• docs/operations.md: running, restarting, logs, troubleshooting
• docs/history.md: what was tried and why it changed course

The living spec, with every measurement, is ~/Documents/Score Exports/PROJECT.md.