360 Secure SkillHub
360 安全技能中心

Suno

围绕 Suno 音乐生成:撰写风格提示词与结构化歌词,可通过托管 API 或浏览器自动化产出音频,并覆盖长曲延长拼接、翻唱、Persona、伴奏与 WAV 导出、商用授权等场景,同时提供故障排查与本地偏好、成功提示词的记忆存储。

ivangdavila
ivangdavila
热度 90.84
最近更新 2026-07-26 20:42:46
分类:创作依赖:内部调用大模型 / 第三方 API / 联网浏览搜索 / 读写文件 / 代码执行应用场景:故事创作 / API 集成 / 自动化脚本 / 工作流

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add suno
Metadata
name
Suno
slug
suno
version
1.0.4
description
Creates music with Suno: crafts style prompts and structured lyrics, generates via API or browser, builds long tracks, covers, and personas. Use when the user wants a song, jingle, anthem, podcast intro, or background track, wants lyrics turned into audio or written for singing, asks to sound like an artist, or when a generation comes out generic, off-genre, cuts off mid-song, sings the style description, garbles words, or gets rejected by moderation. Also covers extending and stitching songs, stems, WAV export, credits, and commercial-use rights. Not for editing or mixing existing audio files.
homepage
https://clawic.com/skills/suno
changelog
Display name shown correctly
metadata
{"clawdbot":{"emoji":"🎵","os":["linux","darwin","win32"],"displayName":"Suno","configPaths":["~/Clawic/data/suno/"]}}

This skill stores all persistent data under ~/Clawic/data/suno/ (config, preferences, project tracking, downloaded audio — see setup.md on first use, memory-template.md for the memory format). If you have data at an old location (~/suno/ or ~/clawic/suno/), move it to ~/Clawic/data/suno/.

Configuration

User-dependent variables. Defaults apply until the user states a preference; store them in ~/Clawic/data/suno/config.yaml.

Variable Type Default Effect
default_method prompts | api | browser prompts Resolves Core Rule 3's default branch: prompts = craft text for manual pasting; api/browser = deliver audio
api_provider aimusicapi | evolink aimusicapi Which hosted service api.md examples target and which env var holds the key
plan_tier free | paid free Gates commercial deliverables (Output Gates), credit budgeting in browser.md, WAV/stems availability in release.md
audio_format mp3 | wav mp3 Download format in release.md; wav requires a paid plan
lyrics_language text (language name) English Language lyrics are written in; when not English, also named in the style field (lyrics.md)

Preference areas to record as the user reveals them:

  • taste — go-to genres, moods, vocal textures, eras; steers first-draft prompts (stated preference → config; observed reactions → memory.md)
  • workflow — present-and-pick vs auto-download both clips; how many rerolls before checking in
  • spend posture — confirm before multi-extend builds or runs beyond a credit threshold
  • conventions — song/file naming, project folder layout under ~/Clawic/data/suno/
  • delivery — where finished audio lands, whether stems are fetched by default

When To Use

  • Generating a song, jingle, theme, intro, or background track with Suno
  • Turning existing lyrics into audio, or writing lyrics meant to be sung
  • Crafting style prompts, tags, or structured lyrics for manual pasting into suno.com
  • Building long tracks, covers, personas, stems, or preparing a commercial release
  • Diagnosing bad output: generic sound, off-genre, garbled vocals, cut-off endings, moderation rejections
  • Not for editing or mixing existing audio files — use audio or ffmpeg

Quick Reference

Situation Play
First run, no ~/Clawic/data/suno/ setup.md
Song cuts off mid-phrase Missing ending cue — [End] as final tag (rule 4)
Style description got sung Words were in the lyrics box — sound in style field, words in lyrics field (rule 1)
Output generic or off-genre Front-load genre, 8-12 precise terms, add era + texture → debug.md
"Make it sound like [artist]" Translate to voice + era + production (rule 6) — names get stripped, credits still spent
Great clip that drifts, or song needs >4 min Extend from the last good timestamp → extend.md
Chorus melody changes every repeat Paste identical chorus text each time (lyrics.md)
New style for an existing song; consistent voice across songs Covers and Personas → covers.md
Jingle, podcast intro, loop, birthday song jingles.md
Stems, WAV, remaster, rights, distribution release.md
Anything else about generating with Suno Reproduce with a minimal prompt in simple mode, then re-add one element per run until it breaks (debug.md)

Depth on demand: debug.md symptom→cause chains · prompts.md style prompt craft · styles.md tag vocabulary · lyrics.md structure, delivery, word budget · extend.md long tracks, crop, stitch · covers.md covers, personas, uploads · jingles.md functional audio briefs · release.md downloads, stems, rights · api.md hosted APIs · browser.md suno.com automation · setup.md first run · memory-template.md memory format.

Core Rules

  1. Style field describes; lyrics field gets sung. Genre words typed into the lyrics box become sung words ("upbeat pop song" turns into the opening line). Sound goes in the style field, words in the lyrics field, structure in [bracketed] tags on their own lines.

  2. Each generation is sampling, not drafting. One run returns two clips from the same prompt — two rolls, not a draft and a revision. When a roll lands, save its exact style string to memory.md and reuse it verbatim; rewording a working prompt resets the odds.

  3. Pick the method by situation.

    Situation Method
    Deliver audio files programmatically Hosted API (api.md)
    No API key, browser tool available Automate suno.com (browser.md)
    User will paste into suno.com themselves Craft prompt + lyrics only
    Unclear (default) default_method from config; if audio delivery is requested and neither key nor browser exists, ask once and record the answer
  4. Force endings. Clips without an ending cue stop mid-phrase. Close the lyrics with an [Outro] section and [End] as the final tag; for instrumentals, [Fade Out] then [End].

  5. Long songs are built, not one-shotted. Generate the strongest opening clip, extend from the last good timestamp (not the raw end), repeat until [End], stitch with Get Whole Song. Target 2-4 minutes; Suno >=3.5 generates 4-minute clips and >=4.5 allows 8-minute songs, but one-shot epics drift in melody and mix. Full workflow and segment planning: extend.md.

  6. Translate artist names into attributes. Moderation rejects or strips artist and brand names — and the credits are still spent. Convert with three knobs: voice texture + era + production. "Like Springsteen" → "raspy heartland rock male vocals, 80s arena production, driving piano and saxophone".

  7. API pattern: generate → poll every 5 seconds → download immediately. Audio URLs expire; generation runs 30-90 seconds. Working code in api.md.

Prompt Essentials

Layered formula (full guide: prompts.md, vocabulary: styles.md):

[genre] [subgenre] [mood] [tempo] [instruments] [vocals] [era/influence]

Example: "indie folk melancholic slow acoustic guitar soft female vocals 90s"

  • Front-load: earlier terms steer harder in practice — genre first, garnish last.
  • Specificity beats quantity: "shoegaze" outsteers "rock, reverb, dreamy, atmospheric".
  • Stay within 8-12 style terms (canonical limits: styles.md).
  • Instrumental: set the instrumental toggle AND leave the lyrics box empty; stray text gets sung.

Output Gates

Before submitting any generation, check:

  • No artist, band, or brand names in prompt or lyrics?
  • Everything that should not be sung is inside [brackets] on its own line?
  • Ending cue present ([End] last) for a standalone song?
  • Style terms 8-12, no contradictions (lo-fi + polished, happy + mournful)?
  • Instrumental toggle matches the request — and lyrics box empty if instrumental?
  • Song over 4 minutes → full lyrics written and segment plan made before the first run (extend.md)?
  • Commercial deliverable → user confirmed a paid Suno plan at generation time (release.md)?

Traps

Trap Why it fails Do instead
Artist/brand names in prompt Rejected or stripped by moderation; credits spent Voice + era + production attributes (rule 6)
Free-plan track for commercial use Rights attach at generation time under the plan then active; free-tier songs are non-commercial under Suno's terms Confirm paid plan before generating deliverables (release.md)
Regenerating a clip that was 90% right Discards the roll that worked Crop the bad part, extend from the last good bar (extend.md)
Lyrics far over 300 words for a 3-minute song Rushed, half-spoken delivery Word budget in lyrics.md
Different chorus wording on each repeat Each variant gets its own melody Paste identical chorus text every time
Negations in the style field ("no drums") Style terms act as attractors; the noun still pulls Omit the term, or use Suno's Exclude Styles field
Prompting an exact duration ("90 seconds long") Duration text is treated as style words, not a control Generate a short structure, then crop to length (jingles.md)
Multi-extend build on an unvalidated idea Burns a day of credits before you know the core works Validate the opening clip first; extend only what already lands (extend.md)

Where Experts Disagree

  • Tags vs natural-language prose. Suno >=4.5 handles prose descriptions as well as comma tags; prose reads better, tags are more reproducible and easier to mutate one term at a time. Default: tags; prose when a mood is easier to describe than to enumerate. Reproducibility needs (albums, clients) → tags.
  • Write lyrics vs let Suno generate them. Generated lyrics are serviceable but generic — competent rhymes, no specifics. Anything with an emotional stake or a client → write them (lyrics.md); throwaway background tracks → generated is fine.
  • One-shot vs extend-built long songs. >=4.5 can one-shot 8 minutes; drift risk rises with length. Boundary: past ~4 minutes, or when any section must be exactly right, build by extending (rule 5).

Security & Privacy

Data storage. This skill creates ~/Clawic/data/suno/ on first use:

  • config.yaml — Declared preferences (variables above)
  • memory file — Observed preferences, successful prompts
  • projects folder — Per-project tracking
  • songs folder — Downloaded audio (optional)

All data stays local. API keys live in environment variables, never in files.

This skill does:

  • Generate music via hosted APIs (requires API key from provider)
  • Navigate suno.com with browser automation
  • Craft optimized prompts for Suno's model
  • Write lyrics with proper structure tags
  • Track projects and successful patterns locally

This skill does NOT:

  • Store API keys in plain text files
  • Access files outside ~/Clawic/data/suno/
  • Make requests without user direction

External endpoints. When using hosted APIs, requests go to:

Endpoint Data Sent Purpose
api.aimusicapi.ai Prompts, lyrics Music generation
api.evolink.ai Prompts, lyrics Music generation
suno.com Browser session Direct platform access

API keys authenticate requests; prompts and lyrics are sent for processing.

Guardrails. By using this skill with APIs, prompts and lyrics are sent to third-party services for music generation. Only use services you trust with your creative content.

Related Skills

More Clawic skills, get them at https://clawic.com/skills/suno (install if the user confirms):

  • audio — Audio processing and editing
  • video — Combine music with video content
  • ffmpeg — Audio format conversion, cropping, cross-fades

Feedback

Part of Clawic, the verified skill library. Get this skill: https://clawic.com/skills/suno.