mm-music-expert（mm 音乐专家）

MiniMax Music Expert

Generate music using MiniMax Music Generation API. Default model: music-2.5+ (recommended). This skill is designed to help music novices create professional-quality music through guided interaction.

Environment Setup

export MINIMAX_MUSIC_API_KEY="your_api_key"

Agent Workflow (MUST FOLLOW)

When a user requests music creation, always follow this 4-step interactive workflow. Do NOT skip steps or jump directly to generation.

Step 1: Understand & Clarify Requirements

Analyze the user's input and identify which dimensions are already covered vs. missing. Then ask targeted follow-up questions to fill the gaps.

Dimension checklist — check which are provided, ask about the rest. Each question MUST provide lettered options (A/B/C/...) for the user to choose from:

Rules:

• 不要一次抛出所有问题。根据用户已提供的信息，只问缺失的关键维度（通常 2-4 个问题）
• 每个问题必须带字母选项(A/B/C/...)，降低用户决策成本
• 选项末尾提供"其他"或"交给你来定"的兜底选项，避免用户被选项限制
• 如果用户的需求已经很明确（覆盖了 Type + Genre + Mood + 至少一个其他维度），可以直接进入 Step 2

示例交互：

用户: "帮我做一首关于杭州的歌" Agent: "好的！关于这首杭州之歌，我还想确认几个方面，你可以直接回复选项字母： 1. 🎵 风格：A. 中国风民谣 B. 流行 C. 古风 D. R&B E. 其他 2. 🎤 情绪：A. 温柔抒情 B. 大气磅礴 C. 欢快明朗 D. 忧伤怀旧 3. 🎶 人声：A. 温柔女声 B. 有力女声 C. 温柔男声 D. 有力男声 4. 🎭 情绪走向：A. 由浅入深慢慢推向高潮 B. 先抑后扬 C. 全程舒缓 D. 起伏交替 5. ✍️ 歌词：A. 我已有歌词 B. 我给主题你来写 C. AI自动生成 直接回复如 1A 2B 3A 4A 5B 即可！"

Step 2: Design Song Blueprint & Present Draft

This is the most important step. Do NOT write a flat prompt. First design a Song Blueprint (段落蓝图) that plans each section's musical characteristics, then derive the prompt and lyrics from it.

Step 2a: Create Song Blueprint

Based on gathered requirements, plan the section-by-section musical journey. Each section should have distinct characteristics.

Blueprint format:

🎼 **歌曲蓝图**

| 段落 | 情绪/能量 | 乐器/编曲 | 人声表现 | 制作特点 |
|------|-----------|-----------|----------|----------|
| [Intro] | ... | ... | ... | ... |
| [Verse 1] | ... | ... | ... | ... |
| [Pre Chorus] | ... | ... | ... | ... |
| [Chorus] | ... | ... | ... | ... |
| ... | ... | ... | ... | ... |

Blueprint example (用户需求: "一首关于杭州的中国风民谣，温柔开头，副歌大气"):

🎼 **歌曲蓝图**

| 段落 | 情绪/能量 | 乐器/编曲 | 人声表现 | 制作特点 |
|------|-----------|-----------|----------|----------|
| [Intro] | 宁静悠远 ⚡① | 古筝独奏琶音 | — | 空间感大，远处水声质感 |
| [Verse 1] | 细腻叙述 ⚡② | +笛子旋律，轻拨琵琶 | 温柔低吟，呼吸感 | 亲密感，少混响 |
| [Verse 2] | 情感渐浓 ⚡③ | +二胡加入，节奏轻启 | 声线渐开，带感情 | 开始加宽声场 |
| [Pre Chorus] | 蓄势待发 ⚡⑤ | 弦乐铺底渐强，鼓点加入 | 力度上升，情绪推动 | 渐强(crescendo) |
| [Chorus] | 大气磅礴 ⚡⑧ | 全编制爆发：弦乐+鼓+笛+琵琶 | 高亢嘹亮，饱满和声 | "音墙"，宽广声场 |
| [Bridge] | 回归内省 ⚡③ | 仅钢琴+古筝 | 轻声吟唱 | 突然安静，对比反差 |
| [Chorus] | 最终爆发 ⚡⑨ | 全编制+合唱层叠 | 全力释放，高音 | 最高能量 |
| [Outro] | 渐归平静 ⚡② | 古筝回归独奏 | — | 渐弱收尾 |

⚡① ~ ⚡⑩ = energy level, 用于可视化段落间的能量变化曲线。

Blueprint design principles:

• 段落间必须有对比：相邻段落的能量/乐器/人声至少有一项明显不同
• 能量曲线要有起伏：避免全程同一能量级。典型曲线：低→中→高→低→最高→低
• Chorus 必须是能量峰值：确保副歌和主歌之间有显著的编曲差异
• Bridge 通常是"呼吸空间"：在两次 Chorus 之间提供反差

Step 2b: Derive Prompt from Blueprint

将蓝图中的信息转化为 prompt。好的 prompt 需要同时描述整体基调和段落间的动态变化。

Prompt construction formula:

[整体风格/BPM] + [人声描述 + 段落间变化] + [核心乐器] + [动态弧线描述] + [声场/制作特点]

Example (based on the 杭州 blueprint above):

A poetic and cinematic Chinese folk pop at 78 BPM, featuring a tender female vocal that transitions from intimate whispering in the verses to powerful soaring belting in the choruses with layered harmonies. The arrangement builds from a solo guzheng intro, gradually adding dizi (bamboo flute), pipa, and erhu through the verses, swelling to a full orchestral "wall of sound" with strings and percussion in the choruses, before retreating to a quiet guzheng outro. The production shifts from a dry, intimate space in the verses to a wide, reverberant arena soundscape in the choruses.

Key: The prompt explicitly describes what changes between sections — not just a flat list of attributes.

Step 2c: Write Lyrics with Section Directions

Write lyrics using annotated structure tags and parenthetical directions to reinforce the blueprint:

[Intro]
（古筝琶音，空灵悠远）

[Verse 1: 温柔低吟]
西湖 烟雨 朦胧了 谁的眼
断桥 残雪 化作 心头的暖

[Verse 2: 情感渐浓]
（笛声加入，轻柔的鼓点）
三潭 映月 点亮了 千年夜
灵隐 钟声 唤醒 梦中的蝶

[Pre Chorus: 情绪递进]
（弦乐渐强 / Rising strings）
这片土地 承载了 多少故事
每一寸 都是 我的根

[Chorus: 大气爆发]
（Full band kicks in / 全编制爆发）
杭州 我的杭州
你是 我心中 永远的歌
...

[Bridge: 回归平静]
（Music drops to piano and guzheng only / 仅剩钢琴与古筝）
闭上眼 你还在那里
...

[Chorus: 最终高潮]
（Maximum energy / 最高能量，加入合唱层）
杭州 我的杭州
...

[Outro: 渐弱收尾]
（古筝独奏，渐弱 / Guzheng solo, fading out）

Step 2d: Present Complete Draft

Combine all above into the final presentation:

📋 **生成方案**

**模式**：Standard Song

🎼 **歌曲蓝图**：
[the blueprint table]

**Prompt**：
[the full prompt — with dynamic arc description]

**歌词**：
[the full lyrics — with annotated tags and parenthetical directions]

**说明**：[1-2 sentences on creative choices]

---
请确认是否满意，或告诉我需要调整的地方（如某段落的情绪、乐器、歌词等）。

Rules:

• Always create the blueprint first, then derive prompt and lyrics from it
• Prompt must describe section transitions, not just a flat style
• Lyrics must use annotated structure tags (e.g., [Chorus: 大气爆发]) and parenthetical directions (e.g., （Full band kicks in）) to control each section's sound
• Show the complete blueprint, prompt, and lyrics — never hide details
• Wait for explicit user confirmation before proceeding to Step 3

Step 3: Generate Music

After user confirms (or after applying their revisions and getting re-confirmation):

1. Determine the output file name from context (e.g., ./audio/杭州之歌.mp3)
2. Run the generation command:

python scripts/generate_music.py \
  --lyrics "<final_lyrics>" \
  --prompt "<final_prompt>" \
  --output ./audio/<filename>.mp3

For instrumental mode, add --is-instrumental true and omit --lyrics. For auto-lyrics mode, add --lyrics-optimizer true and omit --lyrics.

1. Inform user that generation is in progress (API call may take 30-120 seconds)

Step 4: Post-Generation Feedback

After the audio file is generated:

1. Tell the user the file location and invite them to listen
2. Ask for feedback:

🎵 音乐已生成！文件保存在：`./audio/<filename>.mp3`

请试听后告诉我你的感受：
- ✅ 满意，完成
- 🔄 需要调整（请告诉我具体想改什么，比如节奏太快、想换个风格、歌词某处要改等）

1. If user requests changes:

- Analyze what needs modification (prompt, lyrics, or both) - Return to Step 2 with the revised draft - Repeat until user is satisfied

Common revision patterns:

Model & Mode Reference

Model Differences

Generation Modes

Prompt Crafting (Best Practices)

Prompt 控制音乐的整体风格、情绪和制作方向。好的 prompt 不是扁平的属性罗列，而是描述音乐如何在段落间变化。

核心维度（必须覆盖）

进阶维度（提升精度）

Prompt 构建公式（推荐）

[整体风格 + BPM] + [人声描述 + 段落间变化] + [乐器编排 + 段落间叠加/退出] + [动态弧线] + [声场/制作变化]

English template (highest precision):

A [mood] [BPM] [genre] track featuring a [vocal] that [transitions from X in verses to Y in choruses], with arrangement building from [sparse intro instruments] through [verse additions] to [full chorus instrumentation], [bridge contrast], before [final chorus climax]. Production shifts from [verse sound] to [chorus sound].

Example:

A poetic and cinematic Chinese folk pop at 78 BPM, featuring a tender female vocal that transitions from intimate whispering in the verses to powerful soaring belting in the choruses with layered harmonies. The arrangement builds from a solo guzheng intro, gradually adding dizi, pipa, and erhu through the verses, swelling to a full orchestral "wall of sound" with strings and percussion in the choruses, retreating to quiet guzheng in the bridge, before a final explosive chorus with choir layers. Production shifts from dry intimate space in verses to wide reverberant arena soundscape in choruses.

Chinese description template:

78 BPM 的诗意中国风民谣，温柔女声演唱。主歌低吟浅唱、气息感十足，到副歌转为高亢嘹亮的全力释放。编曲从古筝独奏开头，主歌逐渐加入笛子、琵琶，到副歌时弦乐、鼓点全编制爆发形成"音墙"，Bridge回归古筝独奏的宁静，尾声再次爆发后古筝收尾渐弱。

Simple tag template (for straightforward requests):

[风格], [情绪], [BPM], [乐器], [人声类型]

示例: Pop,中国风R&B / Afro House,Tropical House / 流行，女生，舒缓

常见用户需求 → Prompt 转换

Lyrics Writing (Best Practices)

结构标签

官方支持的 14 种标签：

[Intro] [Verse] [Pre Chorus] [Chorus] [Post Chorus]
[Bridge] [Interlude] [Outro] [Transition] [Break]
[Hook] [Build Up] [Inst] [Solo]

结构标签高级用法

编号区分：用编号区分不同段落，使每段有不同表现：

[Verse 1]
街边红灯 摇曳 破了 暮色
[Verse 2]
玉米 粒粒 颗颗 都是 地道

注释标签：在标签后用冒号加注释，指导该段落的演绎方式：

[Intro Hook: Heavy Hanmai Style & 808 Drop]
[Verse 1: Aggressive Rap]
[Pre-Chorus: Build Up]
[Chorus: Maximum Energy]
[Bridge: Comical Trombone Solo]
[Outro: Speed Up]
[Drop / Chorus]

段落对比编排（关键：避免歌曲单调）

核心原则：每个段落的歌词中应通过注释标签 + 括号指令明确告诉模型这一段的音乐特征，确保相邻段落之间产生对比。

完整示例（注意每个段落的括号指令如何控制乐器和能量的变化）：

[Intro]
（古筝独奏琶音，空旷悠远）

[Verse 1: 温柔叙述]
（轻柔的笛子旋律加入，亲密的混音空间）
西湖 烟雨 朦胧了 谁的眼
断桥 残雪 化作 心头的暖

[Pre Chorus: 情绪递进]
（弦乐渐强，鼓点轻启 / Strings building, drums entering）
这片土地 承载了 多少故事

[Chorus: 全力爆发]
（Full band kicks in / 全编制爆发，宽广音场）
杭州 我的杭州
你是我心中 永远的歌

[Bridge: 突然安静]
（Music drops to piano only / 仅剩钢琴，安静的呼吸空间）
闭上眼 你还在那里

[Chorus: 最终高潮]
（Maximum energy, choir layers added / 最高能量，加入合唱）
杭州 我的杭州

对比维度清单（至少变化 1-2 项）：

括号指令（关键技巧）

在歌词中使用 () 或 （） 插入演奏/音效/人声指令，模型会尝试执行：

乐器控制：

[Intro]
（琵琶琶音）
[Outro]
（二胡拉出一个帅气的尾音）
（鼓声重击收尾）

音效/声音设计：

(Sound of glass breaking / 玻璃破碎声)
(Phone ringing sound / 电话铃声)
(Sudden change to 8-bit game sound / 突然变成红白机音效)
(Abrupt silence / 突然安静)
(Sound of a chicken / 一声鸡叫): 咯咯哒？

人声技巧：

(Robotic Voice / 机械音)
(Super fast flow / 极快语速)
(Pitch rising higher and higher / 音调越来越高)
(Deep Voice / 深沉男声)
(Gang Vocals, Everyone shouting)
(Manic laughter)

编曲/动态：

(Heavy Bass kicks in / 重低音爆发)
(Rising synth / 情绪递进)
(Music stops briefly, only Trombone playing funny melody)
(BPM increases significantly to Happy Hardcore speed)
(Fading out with beat / 渐弱)
(Maximum Volume / 最大音量)
(Beat returns - Manyao Style)

拟声词/象声词

在歌词中直接写拟声词可产生特殊音效效果：

Boom Boom Boom!
A-Ba A-Ba A-Ba A-Ba!
Bi-bu Bi-bu Bi-bu Bi-bu (救护车音效)
Womp Womp Womp
嗡——啪!

中文空格分词（节奏控制）

中文歌词中加空格控制字词的演唱节奏和断句：

街边红灯 摇曳 破了 暮色          ← 空格控制断句节奏
半城 烟火 尽是 诱人 醇香

对比无空格版本：

街边红灯摇曳破了暮色              ← 模型自行断句，节奏不可控

双语歌词

中英混合歌词可产生跨文化效果：

霓虹灯在闪烁 (Flash, Flash)
别管现在几点钟 (Tik, Tok)
Zuo! You! Shang! Xia! (左! 右! 上! 下!)

典型歌曲结构模板

标准流行曲：

[Intro]
[Verse 1]
[Pre Chorus]
[Chorus]
[Verse 2]
[Bridge]
[Chorus]
[Outro]

EDM/舞曲：

[Intro]
[Verse 1]
[Pre-Chorus: Build Up]
[Drop / Chorus]
[Verse 2]
[Build Up]
[Drop / Chorus: Maximum Energy]
[Outro]

带 Rap 的歌曲：

[Intro]
[Verse 1: Rap]
[Pre Chorus]
[Chorus]
[Verse 2: Melodic Rap]
[Bridge]
[Chorus]
[Outro]

Script Reference

Key CLI Arguments

API Payload Quick Reference

Standard Song:

{
  "model": "music-2.5+",
  "prompt": "indie folk, melancholic",
  "lyrics": "[Verse]\
街灯微亮晚风轻抚\
影子拉长独自漫步\
[Chorus]\
推开木门香气弥漫",
  "audio_setting": { "sample_rate": 44100, "bitrate": 256000, "format": "mp3" }
}

Pure Music (music-2.5+ only):

{
  "model": "music-2.5+",
  "prompt": "coffee shop, ambient, relaxing, piano",
  "is_instrumental": true
}

Auto Lyrics:

{
  "model": "music-2.5+",
  "prompt": "upbeat pop, summer vibes",
  "lyrics_optimizer": true
}

For full API details, see references/minimax_music_api.md.