Log in
TechniqueBy HypernaturalVersion 4, updated

ElevenLabs music bed

hn/elevenlabs-generate-music

Generate a music bed with ElevenLabs Music: from a prompt and a length (3 to 300 s), or from a composition plan of timed sections (3 to 120 s each, 300 s in all, local and global styles) whose lengths are enforced so a lift or a final hit lands on a cue mark. Instrumental by default, MP3 out. Covers laying the plan's sections on the cut (sparse under the pain, a swell at the turn, a hit on the end card), two contrasting beds, picking the bed under the picked voice, and its level under a voice. Use for a background bed under a voiceover, an ad, a presenter video or motion graphics; the person picks the music.

SKILL.md

Give it a brief (genre, mood, instruments, tempo, the arc and how it ends) and a length, or a plan of timed sections; get an MP3 bed. Use a plan whenever the music must hit a mark: with a prompt alone the timing is only a suggestion. For a spot effect (a whoosh, a door), use elevenlabs-sound-effects; for a song the person brings, there is nothing to generate.

Inputs

Either:

  • A prompt and a length: prompt (genre, mood, instruments, tempo, the arc, how it ends) and music_length_ms, 3 to 300 s. Good for a bed that only has to be roughly long enough. ElevenLabs' API takes up to 600 s, but hn4a refuses more than 300 s, the most ElevenLabs actually makes.
  • A composition plan: composition_plan with sections (each section_name, duration_ms of 3 to 120 s and 300 s in all, lines, positive_local_styles, negative_local_styles), positive_global_styles and negative_global_styles, plus respect_sections_durations: true so the lengths are kept. Lay the sections out on the cut's cue marks (a lift where the product appears, a final hit on the end card).

Run it

The model is elevenlabs/music. Call describe("elevenlabs/music") before the first run in a job and use the field names and limits it gives. To price a batch with run(..., estimate: true), add duration, the length in seconds (30 for a 30000 ms bed): without it the estimate gives no cost.

From a prompt, instrumental:

run("elevenlabs/music",
    {"prompt": "Warm, optimistic indie pop bed, clean electric guitar, soft synth pad, light drums, 100 BPM. Builds gently, lifts at 22 s, ends on a single bright hit with a short ring-out.",
     "music_length_ms": 30000, "force_instrumental": true},
    into="<job>/audio/music/pass_a", unit="bed_a")

From a plan, to land on marks:

run("elevenlabs/music",
    {"composition_plan": {
        "positive_global_styles": ["warm indie pop", "clean electric guitar", "soft synth pad", "100 BPM"],
        "negative_global_styles": ["vocals", "banjo", "slide guitar"],
        "sections": [
          {"section_name": "intro", "duration_ms": 8000, "lines": [],
           "positive_local_styles": ["guitar and pad only, sparse"], "negative_local_styles": ["drums"]},
          {"section_name": "lift", "duration_ms": 16000, "lines": [],
           "positive_local_styles": ["drums enter, fuller, rising"], "negative_local_styles": []},
          {"section_name": "button", "duration_ms": 6000, "lines": [],
           "positive_local_styles": ["one final bright hit, then ring out"], "negative_local_styles": []}]},
     "respect_sections_durations": true},
    into="<job>/audio/music/pass_a", unit="bed_plan_a")
  • force_instrumental goes only with a prompt, never with a plan. Keep a plan instrumental with empty lines in every section and "vocals" in negative_global_styles.
  • Name what must not be there in the negative styles (an instrument, a genre cliché), and check for it anyway (Tips and traps).

Fit the bed to the cut

What worked on ads and explainers cut to a voice:

  • Sections on the cut's marks. Sparse under the pain line, a swell where the turn comes (the brand, the doors opening), full in the middle, one bright hit or a clean button on the end card, no vocals. Write the marks from the shot windows (ffmpeg-cut-clips-to-a-voiceover, --plan) and let references/plan_from_cut.md turn them into the plan: it makes the sections sum to the length to the ms, and merges a section under 3 s into the one before (and says so: move the mark rather than lose it).

    exec python scripts/plan_from_cut.py audio/music/marks.json audio/music/plan_a.json
    
  • Two beds in contrasting styles (warm acoustic and light electronic worked), each "sits under a voiceover" and the cut's length, started with the voice auditions so both land together.

  • Pick the bed under the picked voice, never alone. Played on its own in the first review, the person skipped the music pick; under the chosen read it was easy. If the voice isn't picked yet, offer the beds with the picture and the voice in the same review, or build the live spec once per bed (review_choose).

  • Level: about 0.16 under the voice in a compositor spec, with a fade into the end card; or ducked under the voice in the ffmpeg cut (ffmpeg-cut-clips-to-a-voiceover).

  • Check every take for vocals and for anything the negative styles named with a transcript (whisper-cli-transcribe-with-word-times): words in it mean a voice.

What you get

One MP3 per call. Convert it to 48 kHz WAV before mixing (exec("ffmpeg", ["-i", "audio/music/pass_a/bed_a.mp3", "-ar", "48000", "audio/music/pass_a/bed_a.wav"], cwd)) and read its real length (ffprobe-inspect-media): it isn't always what was asked.

Tips and traps

  • With a prompt alone, timing is a suggestion. One 30 s bed with its marks written in the prompt ("lifts at 22 s, final hit at 29.5 s") landed on them, but other beds faded out 3 to 5 s early, and one ran 37 s for 28 s asked. When a hit must land, use a plan.
  • Listen to every take against the brief. Of two takes asked for "no twang", one had banjo and slide guitar. Put what must not be there in the negative styles, and check each take before it goes to the person.
  • The person picks the music. It is taste, not a measurement: make two or three takes, put them to the person with review_choose, the picture as original, and end your turn; build the mix on the one they pick. This is a gate.
  • The leaf makes the bed; mixing is elsewhere. Lay it under the voice in the compositor spec's audio list at a lower volume (hn4a-compositor), or mix and duck it under the voice with ffmpeg (sidechaincompress keyed on the voice). Normalize the finished mix once (ffmpeg-normalize-loudness), not the bed on its own.
  • Trim a bed that runs long with a short fade-out at the cut's end rather than asking for a new one; regenerate when it ends early or misses a mark.

Check it

  • Its length (ffprobe-inspect-media) against what was asked, and each section mark by ear against its cue in the cut.
  • No vocals, and nothing from the negative styles, over the whole take.
  • Under the voice in the mix: every word stays clear at the bed's level.

Workflows that use this skill

FAQs

Generate a music bed with ElevenLabs Music: from a prompt and a length (3 to 300 s), or from a composition plan of timed sections (3 to 120 s each, 300 s in all, local and global styles) whose lengths are enforced so a lift or a final hit lands on a cue mark. Instrumental by default, MP3 out. Covers laying the plan's sections on the cut (sparse under the pain, a swell at the turn, a hit on the end card), two contrasting beds, picking the bed under the picked voice, and its level under a voice. Use for a background bed under a voiceover, an ad, a presenter video or motion graphics; the person picks the music.

Install Hypernatural Desktop, a free Mac app, and connect it to Claude Code, Codex or Claude Desktop. Then describe the video you want and your agent finds the skill itself, or paste this page's link (https://hypernatural.ai/skill/hn/elevenlabs-generate-music) and it installs hn/elevenlabs-generate-music and follows it.

It runs ElevenLabs music. Hypernatural Desktop has every one of them built in, so there are no API keys to set up, though you can bring your own.

The skill is free. You only pay for the AI generations it runs, and your agent shows you the price of every batch before it spends anything. Editing, captions and titles run on your Mac at no cost.

Yes. Ask your agent to adjust it for your job, or to save its own version. You can publish your version under your own handle so anyone can install it.

Go make something cool.

Install the app, open your agent, describe the video. That's the whole onboarding.

Sticker HypernaturalSticker BannerSticker TVBoy