sumi-sound — Claude Code skill for film audio and setup
sumi-sound is a Claude Code skill that builds a film’s audio after the shot plan is approved. It handles voice, music, and effects, or supports a film with no narration.
MIT ★ 110 updated 2026-11-07
npx skills add beyondtahir/beyondsumi --skill sumi-sound -a claude-codeWhat sumi-sound does
sumi-sound helps shape the audio for a film. Narration can come from Voicebox, ElevenLabs, or a recording supplied by the user; a film can also go without a narrator. ElevenLabs is paid, and each take’s cost needs approval before generation.
The skill selects voices and gets takes approved first. After animation, it adds music and effects, aligns sounds with the shots, and checks whether speech remains clear in the mix. The final audio goes to the renderer with node ../sumi/tools/sumi-render.mjs shots.json --audio sound/mix.wav. That step uses shots.json and sound/mix.wav.
What it helps with
- choose a narration option for a film
- prepare separate voice takes for approval
- sync music and effects with shot actions
- mix audio so dialogue stays audible
Example prompts for the agent
- The shot plan is approved. Choose voices and prepare the film’s narration.
- Add music and effects to the scenes, timed to the shot actions.
- Finish the mix and pass it to the renderer with
shots.json.
How to install sumi-sound
1Skills CLI (Vercel)
npx skills add beyondtahir/beyondsumi --skill sumi-sound -a claude-code2Manually: put the files in place
~/.claude/skills/sumi-sound/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/sumi-sound/
1Skills CLI (Vercel)
npx skills add beyondtahir/beyondsumi --skill sumi-sound -a codex2Manually: put the files in place
.agents/skills/sumi-sound/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add beyondtahir/beyondsumi --skill sumi-sound -a cursor2Manually: put the files in place
.cursor/skills/sumi-sound/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install beyondtahir/beyondsumi skills/sumi-sound2Manually: put the files in place
.github/skills/sumi-sound/SKILL.md1Skills CLI (Vercel)
npx skills add beyondtahir/beyondsumi --skill sumi-sound -a gemini-cli2Platform CLI
gemini skills install https://github.com/beyondtahir/beyondsumi --path skills/sumi-soundManually: put the files in place
.devin/skills/sumi-sound/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about sumi-sound
How do I install sumi-sound in Claude Code?
Run npx skills add beyondtahir/beyondsumi --skill sumi-sound -a claude-code. The skill file is at ~/.claude/skills/sumi-sound/SKILL.md.
What does sumi-sound do?
It helps select narration and prepares music and sound effects for a film. It times audio to the shots and passes the completed mix to the renderer.
What does sumi-sound need to work?
The shot plan must be approved. Narration can use Voicebox, ElevenLabs, or a user-provided recording; ElevenLabs is paid, and each take’s cost needs approval.
Source text
SKILL.md · MIT
--- name: sumi-sound description: "The Beyondsumi sound designer. Choose narration through Voicebox or ElevenLabs, or make a film with no narration. Every film gets library music and sound effects, timed to the action. Use after the shot plan is approved." --- # Sound You design what the audience hears. Silence is one of your tools. A sound should land exactly on the drawing that causes it. ## Narration: ask in this order 1. **Voicebox:** free, runs on the user's computer. See [its companion](../sumi-voicebox/SKILL.md). Check the language first: its engines do not all support every language, and its setup guide does not list Urdu. 2. **ElevenLabs:** paid, with a wide choice of voices. See [its companion](../sumi-elevenlabs/SKILL.md). Approve the cost before each take. 3. **No narration:** the film is carried by captions, music and effects. If the user offers their own recording, use it, following [supplied narration](../sumi/references/supplied-narration.md). Keep the original file unchanged. Save a chosen route only if the user asks you to remember it, in [external memory](../sumi/references/memory.md). ## You work twice when a film has speech 1. **Straight after the Director:** choose the voices, show the cost, then make and approve the takes. The drawings and lip-sync follow these recordings, so redoing a take later can mean paying for drawings again. 2. **After the Animator:** add the music and effects, and make the final mix. ## Every film has music and sound effects Always give a film background music and sound effects, even when it has narration. Build them as the story moves: wind, steps, a ring of steel, a drum on the big moment, and real silence right before it. Keep the voice clear: lower the music under it, and move or quieten any effect that covers a word. Read the finished mix back with a transcription: a line that disappears is buried. ## Expression in a cloned voice - **Do not change the voice.** Keep similarity high, and avoid acting tags on a cloned voice, because they can make it sound like someone else. - **Take phrases, not whole lines.** Generate each phrase separately and use the language's punctuation for feeling: `...` to hold back, `!` to land. - **Put the storytelling gaps in the edit.** Place phrases on the timeline with real silence between them, and time the drawings to those words. - **Clean only:** remove rumble, level each take, and add a little presence. ## Casting voices See [dialogue and casting](references/dialogue-casting.md) when characters speak. More than one voice makes a story more dramatic. The Director can cast a narrator, a second narrating voice and separate characters. - **Give big reactions a sound.** A shocked character gasps or screams, and the sound lands on the drawing where the reaction peaks. Ask for a real reaction from the character's own voice, and check it does not sound like a different person. - **Use an extra voice only for a character who really speaks.** The narrator carries the rest. - Generate each line as its own take, so timing and retakes stay separate. - **Check the language before paying.** Models differ: on ElevenLabs, `eleven_v3` lists Urdu and `eleven_multilingual_v2` does not. Generate one short line first, and read it back with a local transcription if you can. - **Measure each take.** Acting tags and pauses can leave long silence at the end, and tags in the text can spoil the tool's word timings. Time from the sound itself, and trim silence. - **Children:** if the voice tool has no child voice, use a young, bright woman's voice. Never fake a child by speeding up an adult. - Use only voices the user owns or may use, and check they exist on the account first. ## Music and effects Follow [sound routing](../sumi/references/sound-routing.md) and [the sound library](../sumi/references/sound-library.md). - Give every scene a sound plan: music, room sound, and the effects that matter. - Time contacts, like a knock or a footstep, to the exact frame in the shot plan. - See [hooks, endings and dramatic sound](../sumi-director/references/hooks-and-endings.md). ## Final level Aim for about -16 LUFS for YouTube, with peaks below -1 dB. Measure it, and pass the mix to the renderer: ```bash node ../sumi/tools/sumi-render.mjs shots.json --audio sound/mix.wav ``` You cannot judge how something sounds from numbers. Say so, and ask the user to listen.
© beyondtahir · MIT · beyondtahir/beyondsumi