samplette

Sister skill to midi-cook. Cooks MIDI layers that match a specific sample from the Glory MP3 library.

The flow

1. User downloads a sample via the Glory MP3 app → saves to ~/.samplock/samples/ and appends metadata to ~/.samplock/library.json (Glory MP3 auto-detects key + BPM).
2. User asks samplette ("samplette", "pack for [name]", "match my last sample", etc.)
3. Skill reads library.json to get the sample's key + BPM without re-analyzing
4. Skill detects tuning offset from A=440Hz (older records are often -10 to -30 cents flat)
5. Skill drives midi-cook with force_tonic + force_mode + force_tempo_band matching
6. Skill builds a pack folder on Desktop with the original MP3 + N takes of MIDIs

Why a sister skill and not just midi-cook

midi-cook generates fresh random MIDIs for a vibe. samplette guarantees the MIDIs will layer cleanly over a specific sample because the key is forced to match what Glory MP3 detected. Without samplette you'd have to manually figure out the key, manually pass force_tonic/force_mode to cook_random, manually package everything — samplette does it all in one call.

Usage

import sys
sys.path.insert(0, '/Users/<you>/.claude/skills/samplette/scripts')
from samplette import samplette

# Cook a pack for the most recently-downloaded sample
samplette()

# Cook a pack for a specific sample by name (fuzzy-matches titles)
samplette(query="aaliyah")

# Cook more takes
samplette(query="stephanie mills", num_takes=10)

# By full path
samplette(path="/Users/tec/.samplock/samples/Fall In Love.mp3")

Output

~/Desktop/<Sample Title> pack/
├── <Sample Title>.mp3                              ← original sample
├── PACK INFO.md                                    ← key/BPM/tuning offset/take summary
├── Take 1 - <engine> + <voicing> (<BPM> BPM)/
│   ├── chords.mid
│   ├── lead.mid
│   ├── counter.mid
│   ├── melody4.mid
│   ├── 808.mid
│   ├── snare.mid
│   ├── hats.mid
│   ├── open_hats.mid
│   └── info.md
├── Take 2 - ...
└── Take 5 - ...

Tuning offset (A=440Hz)

Old R&B, Japanese city-pop, and a lot of the samples CCBBY digs tend to be off-pitch from the modern A=440Hz standard — usually because they were recorded on tape machines that drifted, or mastered to slightly different reference pitches. AutoKey (the Antares plugin) detects this in CCBBY's workflow; samplette writes the detected cents offset into PACK INFO.md so the producer knows to either (a) pitch-shift the sample onto the grid in FL, or (b) apply the same offset to the cooked MIDIs via their synth's global tuning.

If the offset is >15 cents, it'll sound noticeable. The pack info documents it as, e.g., Tuning: A = 435.8 Hz (-17 cents flat).

Glory MP3 integration

Glory MP3 is a macOS Electron app at /Applications/Glory MP3.app. It downloads YouTube audio and auto-detects key + BPM using its own engine, writing to ~/.samplock/library.json:

{
  "title": "Aaliyah - Rock The Boat (Instrumental)",
  "filePath": "/Users/tec/.samplock/samples/Aaliyah - Rock The Boat (Instrumental).mp3",
  "bpm": 92,
  "key": "G Minor",
  "sourceUrl": "https://www.youtube.com/watch?v=tMbx5ND4RSM",
  "addedAt": "2026-04-23T04:54:06.065Z"
}

samplette reads this file as source-of-truth for key + BPM. No re-analysis needed — Glory MP3 has already done it.

If the user says "download [X] first" or provides a YouTube URL directly, samplette tells them to drop it into Glory MP3 manually (no URL scheme exposed by the app), then re-invoke samplette — it'll pick up the new entry automatically.

Channels CCBBY samples from (for reference, not exhaustive)

• @Musicforemptyrooms (ambient / chill lo-fi)
• @BigPeter1086RealMusicChannel (old R&B rips)
• @freal (samples / loops)
• Plus: Aaliyah, SWV, Destiny's Child, Chief Keef, Japanese R&B/city-pop

Mapping key string → midi-cook params

Glory MP3 emits strings like "G Minor", "C# Minor", "F Major". Parser converts to:

• tonic string matching PITCH_BACK in styles.py (e.g., "G", "C#", "Bb" — normalize A# → Bb, D# → Eb, G# → Ab)
• mode = "minor" or "major" (lowercase)
• tempo_band: slow if BPM ≤ 105, mid if 106-135, fast if ≥ 136. If BPM is unusual (like 199 — probably half-time misdetection), default to slow.

Engine selection for CCBBY's vibe

CCBBY samples OLD R&B heavily. The R&B/soulful data-sourced engines in midi-cook are the best matches:

• rio_leyva — sultry, sparse, patient
• nick_mira — melodic hybrid
• noah_dreamy — dreamy plucky
• taz_og — classic Internet Money melodic trap
• jrhit_melodic — chord-per-beat melodic trap

samplette biases its engine picks toward these 5 over the more aggressive bricc_rage / kc_dark. The close voicing leads (Gold Flint validated).

Error cases

• No library.json or empty → tell user to download a sample via Glory MP3 first
• Query doesn't fuzzy-match → list available samples, ask user to be more specific
• BPM is 0 or missing → default to slow tempo band
• Key isn't parseable → fail loudly with the unparseable string

Future expansion (not MVP)

• Auto-download: take a YouTube URL → write it to Glory MP3's download queue (needs URL scheme or AppleScript automation)
• Sample chord transcription: use engine-fux to also rip the sample's actual chord progression, not just key — layer fresh MIDIs on top of the transcribed chords
• Multi-sample pack: cook in multiple keys at once for sample flips