scenario-ace-step — скилл для Claude Code
Скилл scenario-ace-step для работы с моделями ACE-Step 1.5 на платформе Scenario через MCP: создаёт песни с вокалом из текста, перепевает треки, исправляет отдельные фрагменты, добавляет инструменты и извлекает стемы.
MIT 843 установок ★ 3 964 обновлён 2026-11-01
npx skills add scenario-labs/skills --skill scenario-ace-step -a claude-codeЧто делает scenario-ace-step
Скилл связывает агента с семейством моделей ACE-Step 1.5 на Scenario, разделённым на девять членов по задачам: Text to Music генерирует песню из текста или листа с текстом песни, Cover перепевает или перестилизует существующий трек, Repaint перегенерирует выбранный фрагмент на месте, Add Layer накладывает новый инструмент поверх записи, Complete Track достраивает аранжировку вокруг частичной записи, Stem Extract извлекает одну дорожку из микса. Каждая задача, кроме редакторских, представлена двумя моделями: Turbo для итераций и Quality для финального качества.
Вход — промпт со стилем и настроением, лист с текстом песни размечен тегами [Verse], [Chorus], прописными буквами отмечены ударные фразы, backing-вокал — в скобках. Поля bpm и keyscale задают темп и тональность явно, duration от 10 до 600 секунд управляет длиной трека. Параметр audioCoverStrength в Cover определяет, сколько структуры оригинала сохранится: около 1 оставляет аранжировку узнаваемой, 0.5–0.6 меняет жанр, 0.2 берёт только настроение. Repaint не гарантирует сохранение голоса исходника, поэтому для правки одной строчки с сохранением вокалиста используй Cover с высокой audioCoverStrength, numOutputs до четырёх дублей и ручной выбор подходящего кандидата после прослушивания.
Скилл требует MCP-сервер Scenario с API-ключом в переменной SCENARIO_API_KEY.
Какие задачи решает
- сгенерировать песню с вокалом из промпта и текста песни
- перепеть трек в другом жанре с сохранением структуры
- исправить один куплет или припев без пересоздания всей композиции
- добавить инструментальную дорожку поверх записи с автоподбором темпа и тональности
- извлечь вокал, барабаны или другую дорожку из готового микса
Примеры запросов агенту
- Создай трек с вокалом: стиль — инди-рок, текст песни в файле
lyrics.txt, темп 120 BPM, тональность A Minor, длительность 180 секунд - Перепой трек
audio_asset_idв стиле джаз, сохрани аранжировку на 80%, используй Quality-модель - Исправь припев с 45 по 60 секунду в треке
src_asset, новый текст: «We rise again, through fire and rain»
Как установить scenario-ace-step
1Через Skills CLI (Vercel)
npx skills add scenario-labs/skills --skill scenario-ace-step -a claude-code2Вручную: положить файлы
~/.claude/skills/scenario-ace-step/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/scenario-ace-step/
1Через Skills CLI (Vercel)
npx skills add scenario-labs/skills --skill scenario-ace-step -a codex2Вручную: положить файлы
.agents/skills/scenario-ace-step/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add scenario-labs/skills --skill scenario-ace-step -a cursor2Вручную: положить файлы
.cursor/skills/scenario-ace-step/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install scenario-labs/skills skills/scenario-ace-step2Вручную: положить файлы
.github/skills/scenario-ace-step/SKILL.md1Через Skills CLI (Vercel)
npx skills add scenario-labs/skills --skill scenario-ace-step -a gemini-cli2Командой платформы
gemini skills install https://github.com/scenario-labs/skills --path skills/scenario-ace-stepВручную: положить файлы
.devin/skills/scenario-ace-step/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Через Skills CLI (Vercel)
npx skills add scenario-labs/skills --skill scenario-ace-step -a opencodeClaude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы о scenario-ace-step
Как установить scenario-ace-step в Claude Code?
Выполни npx skills add scenario-labs/skills --skill scenario-ace-step -a claude-code, скилл появится в ~/.claude/skills/scenario-ace-step/SKILL.md.
Что делает скилл scenario-ace-step?
Генерирует музыку с вокалом из текста, перепевает треки в другом стиле, правит отдельные фрагменты композиции, добавляет инструменты и извлекает стемы через модели ACE-Step 1.5 на Scenario.
Как исправить одну строчку в песне без смены голоса?
Используй Cover с исходным треком в srcAudio, исправленным текстом песни в lyrics, audioCoverStrength: 1 и numOutputs: 4. Прослушай дубли, выбери совпадающий по голосу, вырежи нужный фрагмент через model_scenario-audio-cut и вставь его в мастер локально.
Исходный текст
SKILL.md · MIT
---
name: scenario-ace-step
description: "Use when creating or editing music with ACE-Step models on Scenario via MCP: text-to-music songs with sung vocals from a lyric sheet, instrumental tracks, covering or restyling an existing song, repainting one section in place, changing sung lyrics, adding an instrument layer, completing a partial recording into a full arrangement, extracting stems, or choosing between Turbo and Quality lanes. Keywords: ACE-Step 1.5, cover strength, repaint, Vocal2BGM, stem separation, karaoke, BPM."
license: MIT
---
# Scenario ACE-Step Music
## Overview
ACE-Step 1.5, a music production family on Scenario, splits every mode into its own model: Text to
Music, Cover, and Repaint each in a Turbo and a Quality lane, plus a full-quality edit trio (Add
Layer, Complete Track, Stem Extract), nine members at authoring time. Discover them with `search`,
pick the member by mode name, and treat `model_schema_get` as the contract.
Connection and the core loop: see the `scenario` skill in this repo; model-agnostic audio work: the
`scenario-audio` skill. If a sibling skill named here is missing from your available skills, load it
with `skill_get` (Scenario MCP) or ask the user to install it
(https://github.com/scenario-labs/skills#install); unattended, proceed from tool schemas and flag
the gap.
## Quick reference
| Mode | Key inputs | Behavior |
| -------------- | ---------------------------------------------- | -------------------------------------------------------------------------------------------------------------- |
| Text to Music | `prompt` or `lyrics` (either suffices) | full song; `duration` 10 to 600 s (empty lets the model pick), `bpm`, `keyscale` |
| Cover | `srcAudio` (+ `referenceAudio`) | restyles a song, or re-sings it with new `lyrics`; `audioCoverStrength` sets how much of the original survives |
| Repaint | `srcAudio` + window | regenerates `repaintingStart` to `repaintingEnd` in place, length preserved |
| Add Layer | `srcAudio` + `trackName` + `prompt` | adds one stem matched to key, tempo, and groove, optionally windowed |
| Complete Track | `srcAudio` + `completeTrackClasses` + `prompt` | arranges stems around a partial recording (Vocal2BGM) |
| Stem Extract | `srcAudio` + `trackName` | isolates one stem from a mix |
Numbers are at authoring time: read caps off `model_schema_get`. Stem fields take one of twelve
fixed values (vocals, backing_vocals, drums, bass, guitar, keyboard, percussion, strings, synth, fx,
brass, woodwinds), never free text. `srcAudio` is a single audio asset: `upload_asset` first, or
reuse a generated asset id. Across members: `numOutputs` (1 to 4) multiplies cost, `thinking`
(default true, absent on Stem Extract) runs a planning pass, `vocalLanguage`, where present,
defaults to auto-detect. `guidanceScale` exists on the edit trio alone, `audioCoverStrength` on
Cover alone, and `repaintingEnd: -1` runs to the end of the track.
## Lane choice
Turbo and Quality siblings take the same parameters, except `audioFormat` (mp3, wav, flac), at
authoring time on Repaint Turbo alone. No member of the family exposed a `seed`, so two repaints of
the same window are two takes, and the singer can change between them. Repaint regenerates the
window; do not assume it preserves the source voice. Batch with `numOutputs` and pick, rather than
re-running for a match, and read the next section before repainting to fix a sung line. The
asymmetry runs the other way on some section composers in other families, which take a `seed` and no
`numOutputs`; `scenario-audio` carries that rule. Quality is the full-size checkpoint tuned for
fidelity and prompt adherence; Turbo is distilled for speed and at authoring time costs half as
much. Iterate on Turbo (lyrics, structure, windows), then re-run the keeper's parameters on the
Quality sibling. The edit trio has no Turbo lane. `duration` and `numOutputs` both move the price,
so `dry_run=true` before a long track or a batch.
## Writing the song
Keep `prompt` to a one-line style caption (genre, mood, instruments, production) and put the words
in `lyrics`, shaped with section tags: `[Verse]`, `[Chorus]`, direction inside the tag like
`[Bridge - whispered]`, backing vocals in (parentheses), UPPERCASE for belted lines. Vocal-free
music is `instrumental: true`, not "no vocals" in the caption. Tempo and key have dedicated fields
(`bpm`, `keyscale`, e.g. "C Major"), not caption text.
## Editing without regenerating
- Cover keeps the structure recognizable; `audioCoverStrength`: near 1 stays faithful to the
arrangement, 0.5 to 0.6 transforms the genre while the hook survives, around 0.2 borrows only the
mood. `referenceAudio` optionally donates genre and feel; rewritten words go in `lyrics`.
- For a lyric correction where the source voice matters (a flubbed line, a name that must change),
try Cover: `srcAudio` the whole take, `lyrics` the corrected sheet, `audioCoverStrength` at the
default 1, `numOutputs` up to 4 for takes to pick from. High strength favors the source, but does
not guarantee the corrected words, singer, delivery, or timing. Waveform differences alone prove
neither a voice change nor an audible splice. Audition the candidates against the master and
locate the replacement boundaries in both recordings before editing. If listening is unavailable
or no candidate matches, deliver the candidates for review and leave the master unchanged. After
verification, keep only the corrected span: cut it out of the chosen take with
`model_scenario-audio-cut` (a fixed first-party id, Scenario's single deterministic audio trimmer,
so discovery would only re-derive it), `asset_download` it and the master, and splice it in
locally. No MCP tool joins two audio files, so the join is editing work on the downloaded files,
not a gap to bridge with another run; when the track rides a video, the compositor in
`scenario-video-assembly` lays the span over the clip as an audio layer instead.
- To change sung lyrics in a Repaint window when a new take of that section is acceptable: start
`prompt` with "Repaint the selected section with new sung lyrics:", pass the full lyric sheet in
`lyrics` with the new section in place, and set `thinking: false`. Expect the section's voice to
be a new take.
- Complete Track works best on a source with clear melody and rhythm.
## Worked example: fix a chorus in a generated song
1. `search` with `target="models"`, `query="ace-step"`, `public=true`. Pick members by mode name,
e.g. `model_ace-step-1-5-turbo-text-to-music` and `model_ace-step-1-5-turbo-repaint` (live hits
at authoring time: re-discover each session).
2. `model_schema_get` on each id before running it.
3. `model_run` the Text to Music member with `dry_run=true` and
`parameters={"prompt": "indie folk, warm, fingerpicked guitar, soft female vocals", "lyrics": "[Verse]\n...\n[Chorus]\n...", "duration": 90, "vocalLanguage": "en"}`;
re-estimate after changing `duration` or `numOutputs`.
4. Repeat with `wait=false`, then `jobs_wait` with the job id, re-called with `pending_job_ids` on
timeout, never a second `model_run`.
5. `asset_display` to listen; note the chorus window, say 24 to 52 seconds.
6. Decide what the fix must preserve. A new take of the chorus, voice included, is the Repaint
member: `model_run` with
`parameters={"srcAudio": "<asset id>", "repaintingStart": 24, "repaintingEnd": 52, "thinking": false, "prompt": "Repaint the selected section with new sung lyrics:", "lyrics": "<full sheet with the new chorus in place>"}`,
allowing the voice to change. When source voice continuity matters, try the Cover member instead:
`parameters={"srcAudio": "<asset id>", "lyrics": "<full sheet with the new chorus in place>", "audioCoverStrength": 1, "numOutputs": 3}`,
then audition before cutting and splicing as the Editing section says.
7. Listen again, then `asset_download` the keeper (omit `format` for audio).
## Common mistakes
- Expecting one model to infer the mode: every mode is a separate member; pick it by name from
`search`.
- The lyric sheet in `prompt`: it caps at 512 characters and carries style; words go in `lyrics`.
- Free-text stems ("kick drum") in `trackName` or `completeTrackClasses`: only the twelve fixed
values are valid.
- Repainting new sung lyrics without the recipe: the sentinel prompt, the full sheet with the new
section in place, and `thinking: false` go together.
- Repainting to fix one sung line and expecting the same singer back: no seed pins a take and the
window is regenerated, so the voice may change. Cover at full `audioCoverStrength` with corrected
`lyrics` supplies candidates to audition, not a guaranteed performance match.
- Hunting for a `seed` on these members or `numOutputs` on a composer that has a seed instead: batch
here with `numOutputs`; where a member has a `seed` and no batch field, run it several times in
parallel with distinct seeds, each run priced on its own.
- Expecting Repaint to change track length: the window regenerates in place. A longer version of the
same track is the extend lane the `scenario-audio` skill teaches, run by another family's member;
a different length from scratch is a new Text to Music run.
© scenario-labs · MIT · scenario-labs/skills