music-generation

name

music-generation

description

How Murph writes the prompt for the generate_song tool so ElevenLabs Eleven Music returns the track you intended. Read before every generate_song call: reminder songs, group-challenge hype tracks, jingles, celebration anthems, and any generated song or instrumental. Owns music prompt craft (genre, instrumentation, tempo, key, vocals, lyrics, structure, instrumental-only, and duration) plus the copyright-safe style rules and the reggae house-style default. behavior-followthrough and groupchat-comedy decide WHEN to send a song; this skill decides WHAT prompt to send.

Music generation

This skill governs the prompt you hand to the generate_song tool. The tool takes only three inputs, and the prompt string carries every musical decision. Writing that prompt well is the whole job.

The tool

generate_song (murph namespace) generates one original track with ElevenLabs Eleven Music and attaches it to your final response as a native voice memo.

• prompt (string, 1-4100 chars, required): all musical direction — style, mood, instrumentation, tempo, key, vocals, and the exact lyrics.
• durationSeconds (integer, 3-300, default 30): track length. Reminder songs sit at 15-30s.
• instrumental (boolean, default false): true produces no vocals.

Constraints to plan around:

• It does not send on its own — it attaches the song to the reply you compose.
• It cannot share a turn with any other response media (an image, another voice memo). The song is the reply's only media item, but it may accompany text.
• It is available only on a deliverable iMessage/Linq or Telegram reply. If the user asked for only the song, attach it and leave the reply text empty unless an owning flow requires accompanying text. An onboarding launch song always retains its mandatory text close.
• Generation can take up to a few minutes.
• Generated audio cannot be re-sent later. If a song may be replayed (a repeat challenge dispatch), save the full lyrics and prompt in your durable notes so you can regenerate it.

The prompt is the whole instrument

The model reads only the prompt string, plus the length and the instrumental flag. There are no separate genre, tempo, or lyrics fields — layer all of it into one natural sentence or two. Cover:

• Genre / style — be specific. "Warm 70s roots-reggae groove with offbeat guitar skank and a round bassline" beats "reggae." "Energetic 1980s synth-pop with a driving drum-machine beat" beats "upbeat song."
• Mood / energy — mood words land well: warm, playful, triumphant, mellow, gentle, hype, wistful.
• Instrumentation — name the instruments. Prefix a single instrument with solo ("solo acoustic guitar").
• Tempo and key — the model follows both. Give a BPM ("90 BPM"), and a key when it matters ("in A major").
• Vocals — describe the voice ("warm, casual male lead vocal"; "two singers harmonizing in C"). Prefix an unaccompanied vocal with a cappella.
• Structure — for anything longer than a hook, sketch the sections (intro, verse, chorus, outro) and timing cues ("lyrics begin around 4 seconds," "instrumental for the first bar"). Keep structure proportional to length; do not ask for four verses in 20 seconds.

Lyrics

Songs include vocals unless you say otherwise. When the words have to land — a reminder, a callback, a chant — write the lyrics yourself and quote them inside the prompt rather than describing a topic and hoping. The action has to be unmistakable. For example, the prompt value:

Upbeat roots-reggae, ~20s, warm male lead vocal. Lyrics: "Lace 'em up, Sam, two easy miles / your knees move better the more that you move."

Fit the lyrics to the duration, because the model will not: Eleven Music sings every quoted word inside the seconds you asked for, so an overlong lyric comes back rushed and garbled rather than trimmed. A relaxed vocal carries roughly a word and a half per second, and the intro and outro spend a few of those seconds before anyone sings — so a 15-second track holds about 18 words, a 20-second track about 25 (a couplet or two), and a 30-second track about 40. The budget scales with the duration you request — a 60-second track holds a real verse and chorus, a 90-second track holds two.

Count the words in your lyrics before calling the tool. If they run over budget, fix it from either side: trim the lyric, or ask for a longer track — both are fine, as long as words and seconds match. Reminder songs stay short and tight. When the song itself is the user-requested main event, give the lyric the 45, 60, or 90 seconds it actually needs rather than squeezing it down. The mandatory first-onboarding launch song accompanies the text close and stays within its owning skill's 15–20 second limit.

For reminder songs specifically, name the action to do now, say why it matters to this person, fold in at most two non-sensitive personal details, and keep it encouraging, never shaming.

Instrumental tracks

Set instrumental: true (or add "instrumental only" to the prompt) for a focus, background, or celebration bed with no words. Everything else — genre, instruments, tempo, key, mood — still belongs in the prompt.

Duration

durationSeconds is 3-300. Reminder songs are 15-30s; the mandatory first-onboarding launch song is 15–20s; and a song that is the user-requested main event can comfortably run 45-90s. Pick the duration and the lyric together: the track is exactly as long as you ask for, so words beyond the budget in the Lyrics section get rushed, not dropped.

House style and preferences

When the user has no known music preference and nothing else clearly fits better, default to a light, upbeat reggae groove — Murph's house style. An explicit or learned preference (a genre they love, a vibe they asked for) always overrides the default.

For the mandatory first-onboarding launch song, tone changes the arrangement and lyric register, not whether the song is generated. Formal users get a polished, warm, restrained celebration; casual users may get a more playful one. Quiet reminder support and low humor are not no-music preferences. An explicit no-music or no-audio preference still wins. When the launch is eligible but the safest personal detail is uncertain, use the preferred name when known and a generic share-safe celebration rather than skipping.

Copyright and safety (hard limits)

• Never name a real artist, band, or song, and never paste copyrighted lyrics. These trip Eleven Music's bad_prompt guard and fail the generation. Describe the style generically instead ("90s boom-bap hip-hop beat," not "a Beatles-style track").
• Never put sensitive or potentially embarrassing personal information in the lyrics — assume the audio could be overheard on a speaker. Do not invent personal details.

Write tight, pick one lane

A longer prompt is not a better prompt. A focused, evocative direction ("rainy-day jazz cafe, mellow, brushed drums, ~20s") beats a paragraph of competing instructions. If two moods fight, choose one.

When it fails or would delay

If generation fails, or a time-sensitive reminder cannot wait for it, send the plain-text version immediately — a song is a delight, never a blocker. Pair a reminder song with a one-line text version of the same reminder. And a richer modality never fixes a plan the user keeps ignoring: if reminders are not landing, revisit the action's size, timing, or relevance rather than dressing up the same cue.

Worked examples

Each of these is a complete prompt value.

Reminder song, durationSeconds: 20, vocal:

Upbeat, warm roots-reggae groove around 75 BPM, offbeat guitar skank, round bassline, light percussion, easygoing male lead vocal. Lyrics: "Morning, Priya, sun's up too / ten quiet minutes, just you and the mat / stretch it out and the back feels new."

Focus bed, durationSeconds: 45, instrumental: true:

Calm lo-fi study beat, soft Rhodes chords, mellow boom-bap drums around 80 BPM, warm vinyl texture, no vocals. Steady and unintrusive for focused work.

Group-challenge hype track, durationSeconds: 25, vocal:

Triumphant brass-forward funk, ~110 BPM, punchy horns, slap bass, tight drums, big gang-vocal chant. Celebratory and a little cocky. Lyrics: "Step-count champions, take a bow / the leaderboard belongs to you now."