Compose

Compositional Wisdom (from 2,871 canonical masterworks)

These principles were extracted from statistical analysis of Bach, Beethoven, Mozart, Schubert, Chopin, and other canonical composers. Internalize them BEFORE writing any music — they should shape your compositional decisions, not be retrofitted after.

Dynamics are structural, not decorative. Place dynamics at phrase boundaries — 53% of dynamics in canonical music mark structural transitions. Every new phrase should establish its dynamic context. Don't scatter dynamics randomly; use them to articulate form.

Establish character early. Canonical pieces front-load expressive information — 11% of all dynamics appear in the first 10% of the piece. Set the tone immediately. Your opening measures should already have dynamic markings, tempo indication, and expression text.

Don't default to quiet→loud arcs. 55% of canonical pieces have a flat dynamic trajectory. The loudest moment typically falls in the first quarter (median position: 26%), not at a late climax. Local dynamic variation within a steady overall level is more common than a single dramatic build.

Balance crescendo and diminuendo. Canonical music uses nearly equal crescendos and diminuendos (0.92:1 ratio). Don't just build up — tapering down is equally important. A diminuendo after a phrase peak is as expressive as the crescendo that preceded it.

Staccato travels in packs. 87.6% of staccatos in canonical music appear near other staccatos. Use staccato as a passage texture, not as individual note decoration. When a passage is staccato, commit to it for several measures, then release.

Accents highlight the unexpected. 67% of accents in canonical music fall on weak beats. Strong beats already have natural emphasis — accenting them is redundant. Place accents on syncopations, upbeats, and harmonically surprising moments.

Dynamics follow pitch contour. When the melody rises, dynamics tend to increase (46% of cases). When it falls, dynamics tend to decrease (41%). Write dynamics that reinforce the natural shape of your melodic lines.

Use real expression vocabulary. The most common expression markings in canonical scores: dolce, a tempo, rall., ten. (tenuto), rit., sotto voce, una corda, sempre. Use these — they're the language performers understand.

Dynamic range is moderate. Most canonical pieces span pp to ff (not ppp to fff). Reserve extreme dynamics for truly extraordinary moments. A piece that's always at extremes has no room to move.

Phase 2: MusicXML Generation

Generate a complete, valid MusicXML file. Follow these rules strictly:

1. Valid XML structure: Proper <?xml?> declaration, <!DOCTYPE>, <score-partwise> root
2. Part list: Declare all parts in <part-list> before using them
3. Consistent time signatures: Every measure must have the correct number of beats
4. Key signatures: Declare at the start, update on modulations
5. Complete measures: Every <measure> must have notes/rests totaling the correct duration
6. Proper durations: Use <divisions> consistently. Common: divisions=1 means quarter=1, eighth=0.5, etc. Recommend divisions=4 for flexibility (quarter=4, eighth=2, 16th=1)
7. Clefs: Appropriate clef for each instrument
8. Directions: Include tempo, dynamics, and expression markings as <direction> elements

MusicXML Template

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE score-partwise PUBLIC "-//Recordare//DTD MusicXML 4.0 Partwise//EN"
  "http://www.musicxml.org/dtds/partwise.dtd">
<score-partwise version="4.0">
  <work>
    <work-title>TITLE</work-title>
  </work>
  <identification>
    <creator type="composer">Rachmaniclaude AI</creator>
  </identification>
  <part-list>
    <score-part id="P1">
      <part-name>INSTRUMENT</part-name>
    </score-part>
  </part-list>
  <part id="P1">
    <measure number="1">
      <attributes>
        <divisions>4</divisions>
        <key><fifths>0</fifths></key>
        <time><beats>4</beats><beat-type>4</beat-type></time>
        <clef><sign>G</sign><line>2</line></clef>
      </attributes>
      <!-- notes here -->
    </measure>
  </part>
</score-partwise>

Duration Reference (with divisions=4)

Phase 3: Validation & Parse Check

After writing the MusicXML file to output/score.musicxml:

1. Run python -c "from music21 import converter; s = converter.parse('output/score.musicxml'); print(f'Parsed OK: {len(s.parts)} parts, {len(list(s.recurse().getElementsByClass(\"Note\")))} notes')" to verify it parses
2. If parsing fails, fix the XML and retry

Phase 4: Profile Feedback Loop (up to 3 iterations)

Run the feature profile assessment to get ranked improvement instructions:

python -c "
import os, json
from rachmaniclaude.compose.feedback import run_feedback_loop

result = run_feedback_loop(
    musicxml_path='output/score.musicxml',
    output_dir='output',
    classifier_path='models/quality_classifier.joblib' if os.path.exists('models/quality_classifier.joblib') else None,
    regressor_path='models/quality_regressor.joblib' if os.path.exists('models/quality_regressor.joblib') else None,
    distribution_scorer_path='models/distribution_scorer.joblib' if os.path.exists('models/distribution_scorer.joblib') else None,
    profile_path='models/feature_profile.joblib' if os.path.exists('models/feature_profile.joblib') else None,
    quality_threshold=0.5,
    max_iterations=3,
)
print('PASSES:', result['passes'])
print('ITERATION:', result['iteration'])
print()
print(result['critique_text'])
"

Read the profile feedback — it gives you ranked, specific instructions like:

dynamics_count = 0 (percentile 3 in high-rated music, target median: 8). Add dynamic markings throughout the score.

If there are priority improvements to make: Revise the MusicXML to address the top 3-5 issues. Focus on adding expressive markings (dynamics, hairpins, articulations), harmonic variety, and chromatic color — these are typically the highest-impact improvements. Then re-run the assessment to see the delta.

If most features are at or above median: The composition is ready. Stop iterating.

Update output/scratchpad.md after each iteration with what changed and what the delta report shows.

Phase 5: Render Audio

When the feedback loop is done, render the final score:

musescore3 -o output/score.mp3 output/score.musicxml

Output Files

• output/score.musicxml — The generated score
• output/score.mp3 — Rendered audio
• output/scratchpad.md — Composition notes and iteration log
• output/revision_log.jsonl — Feature deltas across iterations

Tell the user the score is ready and offer to play the MP3 or run /assess-quality for additional iterations.