NETCOMPOSER PERFORMER SKILL (plain text, for any LLM or agent)
Version 0.2.0 · CC BY-SA 4.0 · Same content as SKILL.md
Use when: a person sends you to netcomposer.art, or gives you a Composition to perform.

This works with any agent that can read the web and, ideally, run a shell (Grok, ChatGPT, Claude, Gemini,
local agents). It follows https://netcomposer.art/agents/. Published at https://netcomposer.art/create/SKILL.md
(plain text: https://netcomposer.art/create/SKILL.txt) and in the kit repo under skill/.

THE RULES (ALWAYS)

1. Inputs only CC0, CC BY, CC BY-SA or public domain. Never NC, ND, Sampling+ or custom licenses.
   Check each license on the sound's live page or manifest, not from a search result or memory.
2. No pitch-shifting. Time-stretch is fine (duration changes, pitch does not).
3. Measure pitch; do not trust file names or tags. Use a sound for a note only if it measures as
   that note (within about 15 cents).
4. Output license CC BY-SA 4.0. Credit every CC BY and CC BY-SA author by name.
5. Loudness: -16 to -18 LUFS integrated (aim -17), true peak <= -1 dBTP, no clipping.
6. One performance record per Performance (step 5).
7. Call the transforms hacks. Do not publish anything for the person without asking them.

STEP 1. READ THE COMPOSITION

A Composition is Markdown: https://netcomposer.art/composition/ambient-cm.md (current),
https://netcomposer.art/composition/bwv846-harmony-only.md. The list with sha256 is compositions[] in
https://netcomposer.art/netcomposer.json. Someone may also hand you a composition.md made on
https://netcomposer.art/create/ (sections ## Composition and ## Performance).

- The Composition words fix harmony, length and shape. Keep them: the chords, any line the composer
  names, the form. Quote them verbatim in the plan.
- The Performance words are the brief (mood, pacing, density). The rest is your choice: voicings,
  sounds, orchestration, hacks, seed.

STEP 2. WRITE PLAN.JSON

Schema: ncperf/schema/plan.schema.json in the kit. Examples: examples/ambient-cm/perf02.plan.json
and perf03.plan.json. The parts:

- chords: [{t, dur, symbol, voicing:[MIDI]}]. 60 = C4. Prefer voicings that exist in the musicmatter
  full chord set (420 symbols, 2256 voicings; https://fullchords.musicmatter.art/, manifests under
  https://musicmatter.art/sounds/fullchords-v4/<ROOT>/<file>.nc.json, roots A Ab B Bb C D Db E Eb F Fsharp G).
- sources: every sound by id: {"kind":"supplier","manifest":URL}, {"kind":"freesound","id":"123"},
  {"kind":"uiowa_mis","url":URL} or {"kind":"url","url":URL,"license":"CC-BY-4.0","page":URL}.
- banks: instrument -> MIDI note -> candidate ids in order. List two when you can; the renderer keeps
  the first that measures right.
- layers: chord_notes (strike/roll the voicing), chord_pad (time-stretched sustain), chord_under,
  drone, clips (beds, fixed clips; stretch_to_s time-stretches a clip), notes (the line;
  kind:"triplet" for triplets).
- master: per-bus hacks (filters, brighten automation, delay, LUFS match, pan), reverb, tone, loudness -17.
  Reverb: {"engine":"dragonfly","send_lu":5.3} (Dragonfly Hall; send_lu = wet loudness above the dry mix in LU,
  about 5 for ambient, 1-2 for drier; optional params such as decay_s). Or "engine":"convolution","ir":"openair-stpauls-nave"
  for a measured space (CC BY, credited automatically). Tone: high_shelf: {hz, db}.
- quality: stretch_engine (auto = Rubber Band R3 when installed) and resampler (soxr).
- seed, notes (intent and compromises; they go into the record).

Validate: nc interpret --plan plan.json -o checked.json.

STEP 3. RENDER

Getting the kit: nc-performer is public (MIT) at https://github.com/netcomposer-art/nc-performer.
Install: pip install git+https://github.com/netcomposer-art/nc-performer (and ffmpeg with the rubberband filter).
For the default reverb also install pedalboard (pip install pedalboard) and Dragonfly Reverb (VST3; apt install dragonfly-reverb
on Debian/Ubuntu, or the release VST3 in the standard VST3 folder, or point NC_DRAGONFLY_VST3 at it). Both are GPL-3 and optional:
without them the kit warns, uses its synthetic hall, and the record says which reverb ran. rubberband-cli 3.x adds the R3 time-stretch.

With the kit (Python 3.10+, ffmpeg with rubberband, numpy scipy soundfile librosa pyloudnorm):

nc render --plan plan.json -o out/          # fetch, license check, measure, select, hacks, master, record, verify
nc perform composition.md --stub -o out/    # or let the kit draft the plan (no LLM)
nc perform composition.md -o out/           # or any OpenAI-compatible LLM via NC_LLM_BASE_URL / NC_LLM_MODEL / NC_LLM_API_KEY

By hand (any tools), in this order:

1. Fetch each sound. Suppliers first: https://netcomposer.art/suppliers.json -> the Supplier's index ->
   the manifest (license, sha256, audio URL). Check the sha256. Then the open commons (Freesound,
   Wikimedia Commons, Internet Archive, public-domain collections); read the license on the live page.
   Freesound originals need OAuth2; the public HQ previews are fine, say so in the record.
2. Measure every pitched sound (pYIN or similar; FFT for bells). Drop any that is not the note.
3. Place, time-stretch (never pitch-shift), fade, filter, add delay and reverb. Name the reverb tool and its
   version in the record; a convolution impulse response is an input like any other (license, author, sha256).
4. Master to -17 LUFS integrated, true peak <= -1 dBTP: linear gain, then a true-peak limiter (4x oversampled)
   only where a peak would pass the ceiling. Avoid ffmpeg loudnorm: it can switch to dynamic mode on its own. Write FLAC
   24-bit/48 kHz and an MP3 preview.

STEP 4. CHECK

- Loudness and true peak: ffmpeg -i out.flac -af ebur128=peak=true -f null -.
- The chords sound as planned (listen; or python3 tests/check_render.py out/).

STEP 5. THE PERFORMANCE RECORD

JSON, format https://netcomposer.art/record/, example https://netcomposer.art/records/ambient-cm-perf03.json.
It must list every source (URL, license, author, sha256, uses), every setting and the seed, the
interpreter (the model name, or human/agent-written plan), the outputs with sha256 and loudness, credit
lines, and record_hash = sha256 of the canonical JSON (sorted keys, no whitespace) without
record_hash and signature. signature stays UNSIGNED for now. The kit writes it; check with
nc verify-record out/<name>.record.json --audio-dir out/.

STEP 6. SUBMIT

Submissions go through issues on the site repository (or a pull request adding an entry to
suppliers.json). This channel is proposed and may change. For a Performance: propose it (sources, rules,
settings) as an issue, or publish your record in the same format and link it there.
Issues: https://github.com/netcomposer-art/netcomposer.art/issues

Ask the person who sent you before you open the issue or publish anything in their name.