Music Prompt

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)

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 specific model. Valid: suno-v5-5, suno-v4-5, udio-v4, udio-1-5, lyria-3-pro, lyria-3, eleven-music, stable-audio-2-5, stable-audio-open, musicgen, tencent-song-generation, sonauto-v2, riffusion, mubert
• music-prompt <brief> --genre <genre> — force a genre recipe. Valid: hyperpop, drill, country, lo-fi, ambient, orchestral, k-pop, afrobeats, jazz-fusion, hardcore-punk, synthwave, gospel, plus free-form
• music-prompt <brief> --instrumental — instrumental only; lyrics box empty or omitted
• music-prompt <brief> --lyrics file:<path> — bring your own lyrics; skill embeds tags into them
• music-prompt <brief> --exclude "<styles>" — exclude-styles (works on ElevenLabs Music; degrades to negative-style note on others)
• music-prompt <brief> --cover <reference-track-description> — Suno Cover or Udio remix mode
• music-prompt <brief> --extend — Suno Extend or Udio extension mode; prompt continues an existing track
• music-prompt <brief> --variants 3 — 3 alternatives with different tag stacks
• music-prompt <brief> --improve — user gives a weak prompt + describes the bad output; skill rewrites
• music-prompt <brief> --execute — also call the API if env var for --model is set; save MP3 to ./generated/music/
• music-prompt <brief> --execute --output <dir> — custom output dir
• music-prompt <brief> --execute --yes — skip cost confirmation
• music-prompt --check --model <slug> — verify env + connectivity, no generation
• music-prompt --list-providers — list executable providers given current env (music modality)

REFERENCES (load on demand)

EXAMPLES

See examples/before-after.md — 6 calibration pairs covering anthemic modern pop (Suno), drill verse (Suno), long-form jazz fusion (Udio), label-safe orchestral (Lyria 3 Pro), indie folk with exclude-styles (ElevenLabs), RU pop ballad with language switching at section boundaries (Suno).

CONSTRAINTS

• Suno: brackets go in the LYRICS box ONLY. The Style of Music box accepts natural language only — brackets there are ignored or break the gen.
• Don't invent tags. Every bracket tag must come from the canonical taxonomy in references/meta-tags.md OR its deep-dive extensions: vocal-tags.md (voice character / register / effects sub-tags), instrumental-tags.md (per-family instrument sub-tags), mix-production-tags.md (era / reverb / delay / saturation sub-tags). Together these four files form the full catalog. Tags outside this catalog ([Dubstep Drop] if not on the list) are ignored or hurt.
• Max 4-8 tags per stack. Past 8, the model dilutes. Front-load core genre + era + mood; instruments + FX + vocal direction follow.
• One tag per category in a stack. Don't combine two genres in one stack ([hyperpop | country]) unless deliberate fusion — combine two genres at most.
• Order matters: core genre → era → mood → instrument → mix/FX → vocal direction.
• Section headers on their own line. ALWAYS. [Chorus] then newline then lyric. Inline [Chorus] We light it up is sung literally.
• Ad-libs in parens on own line, ≤3 words. (yeah), (uh, ad-lib: oh).
• Repetition: duplicate the line. (x2) is unreliable across models.
• Language switch at SECTION boundary, not within a line. Verse 1 RU, Chorus EN is fine. Я walking down the street tonight is not.
• Don't name living artists or copyrighted songs. Suno / Udio / Lyria scrub or refuse. Use era + scene descriptors instead (80s glam metal, 2010s EDM festival, Y2K R&B).
• Don't write stage directions outside brackets. "Then she shouts:" is sung as a lyric. Use [Verse | shouted vocal] instead.
• Don't mix conflicting moods. [dark] + [happy] randomizes the output. Pick one emotional anchor per section.
• Don't bury the genre in the Style box past char-1000 on Suno. Silent truncation. Front-load.
• Lyria 3 Pro refuses artist-mimicry prompts. Don't write "in the style of Hans Zimmer". Use "modern epic film score, brass swell, taiko drums" instead.
• Stable Audio 2.5 is weak on vocals. Treat as composition / sound design — don't expect singing.
• For two-box models (Suno, Udio): output MUST clearly separate Style and Lyrics blocks. Don't paste them as one prompt.
• --execute is opt-in. Default flow stays prompt-only (two-box output for Suno/Udio, single prompt for others).
• Never print API keys. Mask in errors. Refer to env var names, not values.
• Confirm cost. Suno / Lyria / Eleven are near or above $0.10; user must answer Y unless --yes.
• Output dir is ./generated/music/ by default. Don't write outside it without explicit --output.
• API failure → fall back gracefully. Save the Style box + Lyrics box (or single prompt) to ./generated/music/<timestamp>-prompt-only.txt with reason — the user can still paste manually.
• Suno two-box → CLI mapping (for --execute): Style box → --prompt; Lyrics box → --lyrics-file <path> or --lyrics "...". --instrumental empties the lyrics field.

INVOCATION HINTS

When the user says any of:

• "prompt for Suno / Udio / Lyria / Eleven Music / Stable Audio / MusicGen / Riffusion / Mubert / Sonauto"
• "song about X", "track about X", "make me a song / track / instrumental"
• "AI music prompt", "music meta tags", "Suno meta tags"
• "cover song", "remix this track", "extend this song"
• "background loop", "jingle", "stinger", "intro music"
• "orchestral cue", "film score cue", "trailer music"
• "hip-hop verse / drill verse / lo-fi loop / synthwave instrumental / gospel chorus"
• "improve this music prompt"
• "execute the music prompt", "actually render the song", "fire the gen on Suno / Eleven / Lyria"
• "save the track", "render the song"

RU triggers:

• «промпт для Suno / Udio / Lyria / Eleven Music / Stable Audio / MusicGen»
• «песня про X», «трек про X», «инструменталка про X»
• «AI-музыка промпт», «meta-теги для Suno»
• «cover-версия», «ремикс этого трека», «продли эту песню»
• «фоновая музыка», «джингл», «стингер», «вступление»
• «оркестровая кю / трейлер-музыка»
• «лофай-луп / синтвейв-инстру / госпел-припев»
• «улучши промпт для музыки»
• «выполни music-промпт», «отрендери трек», «вызови Suno / Eleven / Lyria», «сохрани песню»

The Style box and Lyrics box are usually written in English (Suno / Udio / Eleven parse EN best). Lyrics can be RU / multilingual where the model supports it (Suno + Eleven handle multilingual; Lyria limited to EN/ES/FR/JP; Udio works but less stable). Switch language at SECTION boundary, never within a line.

For pure-instrumental loops or background music: use --instrumental mode; lyrics box empty.

Use this skill. For static image — image-prompt. For video — video-prompt. For in-clip dialogue audio (4-15s talking-head with lip-sync) — video-prompt with --audio mode, not this skill.