/music-prompt

name: music-prompt
description: "Write prompts for 10+ frontier AI music generators (Suno v5.5, Udio v4, Lyria 3 Pro, ElevenLabs Music, Stable Audio, MusicGen, Sonauto, Riffusion, Mubert). 2026 meta-tag taxonomy, Style+Lyrics workflow, exclude-styles. Use when: 'prompt for Suno / Udio / Lyria / Eleven Music', 'song about X', 'meta tags', 'cover song style'."

license: MIT
allowed-tools:
  - Read
  - Write
  - Edit
  - Bash
  - Grep
  - Glob

<objective> Generate a tight, model-aware prompt for AI music generation. Output: for two-box models (Suno, Udio) — separate Style of Music box + Lyrics box with embedded bracketed meta-tags; for single-prompt models (Eleven, Lyria, Stable Audio) — one well-formed prompt with bracketed cues and timing markers; for field-driven models (Lyria) — natural-language prompt + structured key/BPM/lyrics fields.

This skill does NOT call the music model — it produces text you paste into Suno / Udio / Lyria 3 / ElevenLabs Music / Stable Audio / Riffusion / Mubert / MusicGen / SongGeneration / Sonauto.

Use when the user wants a song, instrumental cue, cover, remix, jingle, background loop, talkbox / vocal arrangement, or a specific genre recipe. The skill picks the right model + correct tag taxonomy + the box-placement rules per model, and produces paste-ready output instead of generic "AI music".

This skill does NOT:

• generate the audio itself (that's the model)
• write traditional sheet music or MIDI files
• handle voice cloning workflow (that's a Suno Voices / ElevenLabs feature inside their tool)
• generate spoken-word audio for video/podcasts (use `video-prompt` audio block for in-clip dialogue; for standalone speech, use ElevenLabs voice tools)

</objective>

ROLE

Read the request → identify subject + genre + intent (full song / instrumental / cover / loop / cue) → pick target model from `references/model-picker.md` → assemble Style box + Lyrics box (or single prompt depending on model) using the 8-category meta-tag taxonomy → apply per-model placement rules → return paste-ready output.

PIPELINE

1. **Clarify if needed.** If the user gave only a topic ("song about loneliness"), pick a sensible genre default (modern pop ballad) and check before committing. If they gave genre + mood + structure, skip.

2. **Pick model from intent** — see `references/model-picker.md`:

• Full song with vocals + lyrics → Suno v5.5
• Longest coherent song (5-10 min) → Udio v4
• Cinematic / orchestral cue → Stable Audio 2.5 or Lyria 3 Pro
• Sound-design / instrumental → Stable Audio 2.5
• Cleanest licensing → Lyria 3 Pro or ElevenLabs Music
• Exclude-styles control → ElevenLabs Music
• Stems / multi-track output → Suno (12 stems) or Udio (Stem Sep 2.0) or Sonauto v2
• Voice clone → Suno Voices (Premier)
• Cover / remix → Suno Cover or Udio remix
• Self-host / open-weights → MusicGen / Stable Audio Open / Tencent SongGeneration
• Background loops for apps → Mubert API
• Free unlimited iteration → Riffusion

3. **Pick the genre recipe** — see `references/genre-recipes.md` for 12 ready-made stacks (hyperpop / drill / country / lo-fi / ambient / orchestral / K-Pop / Afrobeats / jazz fusion / hardcore punk / synthwave / gospel). Adapt; don't invent fresh stacks unless explicitly creative-fusion.

4. **Build the prompt** using the model's correct workflow:

• **Suno (two-box)**: Style of Music box (≤1000 chars, natural-language genre/mood/instrument/era; NO brackets), Lyrics box (≤3000 chars, all `[bracketed]` section + delivery + FX tags on their own lines + the lyric content). See `references/models/suno.md`.
• **Udio (one prompt + lyrics)**: style prompt natural-language; lyrics with `/`-inserted bracketed structural tags (atomic, no `|` stacking). See `references/models/udio.md`.
• **ElevenLabs Music (single prompt)**: natural language with bracketed style cues + timing markers + exclude-styles. See `references/models/elevenlabs.md`.
• **Lyria 3 Pro (field-driven)**: natural-language prompt + separate `lyrics`, `key`, `BPM` fields. No brackets. See `references/models/google.md`.
• **Stable Audio 2.5 (free text)**: instrumental-focused, free-text prompt with multi-part composition. See `references/models/stable-audio.md`.

5. **Apply tag taxonomy** — see `references/meta-tags.md` for the canonical 8-category list (Structure / Vocal delivery / Vocal effects / Instrumental / Mix-production / Energy-dynamics / Era-genre / FX). Max 4-8 tags per stack. Order: core genre → era → mood → instrument → mix/FX → vocal direction. One tag per category in a stack.

6. **Apply lyrics conventions** — see `references/lyrics-conventions.md`:

• Section headers on their own line.
• Ad-libs in parens on their own line, ≤3 words.
• Repetition: physically duplicate the line; `(x2)` unreliable.
• Language switch: at section boundary, not within a line.
• Stage directions ONLY in brackets; never as bare prose.

7. **Output.**

• For Suno / Udio: two fenced blocks clearly labeled `Style of Music` and `Lyrics`.
• For single-prompt models: one fenced block.
• For field-driven (Lyria): structured `prompt:` / `lyrics:` / `key:` / `BPM:` block.
• 1-line note: which model + recipe + key conventions applied.
• If `--variants N` requested — N alternatives with different tag stacks or genre lean.

8. **(Optional) Execute via API.** If `--execute` was passed AND the env var for the chosen model is set, also run `python3 scripts/run.py --model <model> --prompt <style-box-text> --lyrics-file <lyrics.txt>`. The runner polls the vendor and saves an MP3 to `./generated/music/`. Cost confirmation prompts (Suno / Lyria / Eleven are near or above $0.10) unless `--yes`. On any failure, fall back to prompt-only — the Style box + Lyrics box stay paste-ready. See `references/execute.md`.

MODES

• `music-prompt <topic-or-brief>` — generate default prompt (intent-routed model)
• `music-prompt <brief> --model <name>` — target a speci