lyric-diagnostic

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Lyric Diagnostic: Skill

歌词诊断Skill

You diagnose lyric problems across multiple dimensions. Your role is to identify why lyrics fail and guide writers (human or AI) toward more effective choices.

你需要从多个维度诊断歌词问题，职责是找出歌词欠佳的原因，并引导创作者（人类或AI）做出更有效的选择。

Core Principle

核心原则

Lyrics fail in predictable, diagnosable ways. Generic is not neutral—generic is bad.

AI-generated lyrics cluster at statistical distribution centers: the most probable word choices, rhyme pairs, and emotional vocabulary. Probability optimization produces mediocrity. Effective lyrics require intentional deviation from defaults.

The same pattern that makes "Celtic" trigger tinwhistle in music generation makes "love" trigger predictable rhyme pairs and emotional cliches in lyric generation.

歌词的失败存在可预测、可诊断的模式。平庸绝非中性——平庸就是差。

AI生成的歌词集中在统计分布的核心区域：最可能的用词、押韵组合和情感词汇。概率优化只会催生平庸之作。优质歌词需要刻意偏离默认选项。

就像“Celtic”会在音乐生成中触发锡哨声一样，“love”这类词会在歌词生成中触发可预测的押韵组合和情感陈词滥调。

The Lyric States

歌词问题状态

State LY1: Cliche Trap

状态LY1：陈词滥调陷阱

Symptoms: Predictable rhyme pairs dominate. Word choices are first-associations. Nothing surprises. Lyrics could fit any song on this topic.

Key Questions:

What rhyme pairs are you using? Are they first-associations?
Would replacing the end-rhyme word break the meaning? (If not, it's filler)
Can you enumerate the default words for this topic and avoid them?
What would a different genre use for this same emotion?

Diagnostic Checklist:

No Tier 1 cliche words (love, heart, baby, forever)
No stock phrases (see list below)
Rhyme pairs are not obvious first-associations
At least one surprising word choice per verse
Specific details rather than generic statements

Stock Phrases (phrase-level cliches, not just words):

Scripts can't catch these—they require semantic judgment:

Phrase	Category	Why It's Cliche
"screaming inside"	unexpressed anguish	Every song about hidden pain uses this
"need you gone"	direct desire	States want instead of showing it
"won't leave me alone"	persistence	Generic memory/thought complaint
"thought I died"	emotional death metaphor	Overused intensity marker
"used to be" + past state	nostalgia setup	Stock before/after framing
"look in your eyes"	connection	Vague romantic gesture
"torn apart"	destruction	Abstract violence metaphor
"can't let go"	attachment	States the problem directly

Lazy Rhyme Pairs (rhyme drove word choice):

Pair	Problem
desire / fire / tired	Cluster appears together constantly
inside / died / alive	Death-interior cluster
away / stay / day	Motion-time cluster
heart / apart / start	Heart-destruction cluster
alone / gone / on	Isolation cluster

Interventions:

Run
```
cliche-check.ts
```
as pre-filter (surfaces candidates, doesn't diagnose)
Scan for stock phrases manually—scripts miss these
List the 10 most predictable rhymes for your end words—avoid them all
Replace abstract emotions with concrete sensory imagery
Ask: "What would someone who hates this genre never say?" Use that.

症状： 可预测的押韵组合占主导。用词是第一联想词。毫无惊喜感。歌词可适配同主题的任意歌曲。

核心问题：

你使用了哪些押韵组合？是否是第一联想词？
替换结尾押韵词会破坏表意吗？（如果不会，那就是填充词）
你能否列举该主题的默认用词并避开它们？
其他流派会用什么词汇表达同样的情感？

诊断检查清单：

无一级陈词滥调词汇（love、heart、baby、forever）
无套话（见下方列表）
押韵组合并非明显的第一联想词
每段主歌至少有一个令人惊喜的用词
使用具体细节而非笼统表述

套话（短语级陈词滥调，不止是单个词）：

脚本无法识别这些——需要语义判断：

短语	分类	为何是陈词滥调
"内心嘶吼"	未表达的痛苦	每一首关于隐藏痛苦的歌都在用这个
"需要你离开"	直接诉求	只说想要什么而非通过场景展现
"不肯放过我"	执念	关于回忆/思绪的笼统抱怨
"感觉自己已死去"	情感死亡隐喻	被过度使用的强度标记
"曾经"+过往状态	怀旧铺垫	常见的前后对比框架
"凝视你的眼眸"	情感联结	模糊的浪漫表述
"支离破碎"	情感崩塌	抽象的暴力隐喻
"无法释怀"	情感依附	直接点明问题而非展现

"偷懒式"押韵组合（为押韵而选词）：

组合	问题
desire / fire / tired	这些词总是同时出现
inside / died / alive	死亡-内心类集群
away / stay / day	动作-时间类集群
heart / apart / start	心脏-破碎类集群
alone / gone / on	孤独类集群

干预措施：

运行
```
cliche-check.ts
```
作为预筛选工具（仅提供候选，不做诊断）
手动排查套话——脚本无法识别这些
列出结尾词的10个最可预测押韵词，全部避开
用具体的感官意象替换抽象情感
思考：“讨厌这个流派的人绝对不会说什么？”然后用上这个表达

State LY2: Meter Mismatch

状态LY2：韵律不匹配

Symptoms: Lines don't fit the melody. Stresses fall on wrong syllables. Unnatural pronunciation required. Syllables crammed or stretched awkwardly.

Key Questions:

Can you tap the rhythm and speak the line simultaneously?
Where do stresses fall? Do they match the melody's emphasis?
Are any syllables swallowed or stretched unnaturally?
How many syllables per phrase? Is there consistency?

Diagnostic Checklist:

Natural word stress aligns with musical beats
Syllable counts fit the melodic phrase
No awkward pronunciations required
Contractions/elisions are singable

Interventions:

Run
```
meter-check.ts
```
to analyze syllable counts and stress patterns
Speak the lyrics while tapping the beat—mismatches become obvious
Swap words with same meaning but different syllable count
Use contractions to reduce syllables, or expand to add them

症状： 歌词与旋律不契合。重音落在错误音节上。需要刻意扭曲发音。音节被生硬挤压或拉长。

核心问题：

你能否一边打节拍一边念出歌词？
重音落在何处？是否与旋律的强调点匹配？
有没有音节被吞掉或刻意拉长？
每个短语有多少音节？是否保持一致？

诊断检查清单：

自然的单词重音与音乐节拍对齐
音节数适配旋律短语
无需刻意扭曲发音
缩写/省音适合演唱

干预措施：

运行
```
meter-check.ts
```
分析音节数和重音模式
打节拍的同时念歌词——不匹配的地方会很明显
替换成表意相同但音节数不同的词
使用缩写减少音节，或扩展表述增加音节

State LY3: Shallow Surface

状态LY3：表意肤浅

Symptoms: Lyrics say exactly what they mean and nothing more. No subtext, no layers, no depth. Emotions are stated directly. Everything is on the surface.

Key Questions:

Is there a second reading? Something between the lines?
Are you stating the emotion or evoking it through imagery?
What specific concrete image grounds the abstraction?
Could this be understood on first listen but reveal depth on repeat?

Diagnostic Checklist:

At least one image per verse that works on multiple levels
Emotions evoked through detail, not stated directly
Something for repeat listens to discover
Concrete specifics anchor abstract concepts

Interventions:

For every abstract statement, find a concrete action/image that embodies it
Instead of "I'm sad," show what sad looks like, sounds like, feels like
Add a detail that doesn't quite fit—creates productive ambiguity
Test: can you explain two different interpretations of the same line?

症状： 歌词直白表述意图，毫无深层含义。无潜台词、无层次、无深度。直接点明情感，所有内容浮于表面。

核心问题：

是否有第二种解读方式？是否有言外之意？
你是直接点明情感还是通过意象唤起情感？
用什么具体的意象支撑抽象概念？
初次聆听就能理解，但重复聆听能发现深层内容吗？

诊断检查清单：

每段主歌至少有一个具有多层含义的意象
通过细节唤起情感而非直接点明
有可供重复聆听挖掘的内容
用具体细节锚定抽象概念

干预措施：

每一个抽象表述都对应一个具体的动作/意象
不说“我很悲伤”，而是展现悲伤的样子、声音、触感
添加一个看似不相关的细节——制造有益的模糊感
测试：你能否对同一句歌词给出两种不同的解读？

State LY4: Register Mismatch

状态LY4：语体不匹配

Symptoms: Lyric tone contradicts the music's emotion. Casual words over dramatic instrumentation. Heavy lyrics over light music. Vocabulary doesn't match genre expectations.

Key Questions:

What emotion should the listener feel at this moment?
Is your vocabulary elevated/casual matching the music?
Does the distance (intimate/observational) match the instrumental?
Would this language make sense in this genre?

Diagnostic Checklist:

Vocabulary matches musical intensity
Distance (I/you vs. they/we) fits the sound
Genre vocabulary expectations met or intentionally broken
Irony (if present) is signaled clearly

Interventions:

Listen to the music without lyrics—what emotions does it evoke?
Match lyric intensity to musical intensity section by section
Check genre vocabulary norms—violate intentionally or not at all
If using irony, add signal words or structural cues

症状： 歌词语气与音乐情感矛盾。随意用词搭配戏剧性编曲。沉重歌词搭配轻快音乐。词汇不符合流派预期。

核心问题：

此时听众应该感受到什么情绪？
你的词汇正式/随意程度是否与音乐匹配？
距离感（第一/第二人称 vs 第三/第一人称复数）是否与编曲契合？
这种语言在该流派中合理吗？

诊断检查清单：

词汇强度与音乐强度匹配
距离感（我/你 vs 他们/我们）与音效契合
符合流派词汇预期或刻意打破预期
讽刺（如果使用）有明确信号

干预措施：

不听歌词只听音乐——它唤起了什么情绪？
逐段调整歌词强度以匹配音乐强度
核对流派词汇规范——要么刻意打破要么完全遵循
如果使用讽刺，添加提示词或结构线索

State LY5: Sound Neglect

状态LY5：发音忽视

Symptoms: No attention to how words feel in the mouth. Harsh consonant clusters at tender moments. Monotonous vowel sounds. No alliteration, assonance, or deliberate sound texture.

Key Questions:

What consonant sounds dominate? What texture do they create?
Do vowel sounds support the emotional content?
Are there harsh clusters that disrupt flow?
Is there intentional sound painting?

Diagnostic Checklist:

Soft consonants (l, m, n, w) at intimate moments
Hard consonants (k, t, p) at impactful moments
Vowel sounds support emotional content
At least one sound device per verse (alliteration, assonance)

Interventions:

Read lyrics aloud focusing only on how they feel in the mouth
Map emotional moments to appropriate sound textures
Replace harsh-sounding words at tender moments
Add deliberate alliteration or assonance at key phrases

症状： 不关注词语的发音感受。温柔时刻出现刺耳的辅音集群。元音音单调乏味。无头韵、腹韵或刻意的发音质感。

核心问题：

主导的辅音是什么？它们营造了什么质感？
元音是否支撑情感表达？
是否有刺耳的集群破坏流畅性？
是否有刻意的音画同步？

诊断检查清单：

温柔时刻使用软辅音（l、m、n、w）
冲击性时刻使用硬辅音（k、t、p）
元音支撑情感表达
每段主歌至少使用一种发音手法（头韵、腹韵）

干预措施：

大声朗读歌词，只关注发音感受
将情感时刻对应到合适的发音质感
替换温柔时刻的刺耳词汇
在关键短语处刻意添加头韵或腹韵

State LY6: Structural Chaos

状态LY6：结构混乱

Symptoms: No identifiable hook. Verse and chorus feel interchangeable. Repetition is annoying, not reinforcing. Bridge doesn't contrast. Song has no arc.

Key Questions:

Is there a clear hook? Where does it land?
Does each verse add new information or repeat?
Does the bridge actually contrast or just continue?
Is repetition earning its place?

Diagnostic Checklist:

Hook is identifiable and memorable
Verses differentiate (new info, different angle)
Chorus delivers emotional payoff
Bridge provides contrast or escalation
Repetition evolves or emphasizes, doesn't bore

Structural Cliches:

Pattern	Problem	Fix
Bridge = Verse	Not a bridge at all—just more of the same	Shift perspective, time, or speaker
Chorus expands on repeat	Adding lines without development	Each chorus iteration should mean more
Hook repeated without exploration	Says the phrase but never unpacks it	What does the hook mean? Show us.
No arc	Emotional state same at end as beginning	Something must change—realization, acceptance, escalation
Verses say same thing differently	Verse 2 is Verse 1 with synonyms	Each verse needs new information or angle

Interventions:

Identify the hook—if you can't, create one
Make each verse answer a different question about the topic
Bridge should shift perspective, time, or intensity
Cut repetition that doesn't earn its place
Check: is the speaker in a different place emotionally at the end?

症状： 无明确钩子。主歌和副歌可互换。重复内容令人厌烦而非强化记忆。桥段无对比。歌曲无情感弧线。

核心问题：

是否有清晰的钩子？位置在哪里？
每段主歌是否添加新信息还是重复内容？
桥段是否真的形成对比还是只是延续？
重复内容是否有存在的价值？

诊断检查清单：

钩子清晰且令人难忘
主歌各有不同（新信息、不同角度）
副歌传递情感高潮
桥段提供对比或升级
重复内容是演化或强调，而非冗余

结构类陈词滥调：

模式	问题	修正方案
桥段=主歌	根本不是桥段——只是重复内容	转换视角、时间或叙述者
副歌重复时扩展内容	添加内容却无发展	每次副歌重复都应承载更多含义
钩子重复却无深入挖掘	只说短语却不展开	钩子的真正含义是什么？用场景展现
无情感弧线	结尾与开头的情感状态一致	必须有变化——顿悟、接受、升级
主歌换词不换意	第二段主歌只是第一段的同义词替换	每段主歌需要新信息或新角度

干预措施：

识别钩子——如果找不到，就创造一个
让每段主歌回答关于主题的不同问题
桥段应转换视角、时间或强度
删除无价值的重复内容
检查：歌曲结尾时叙述者的情感状态是否有变化？

Diagnostic Process

诊断流程

Step 1: First Listen (Holistic)

步骤1：初次聆听（整体感知）

Listen without analyzing. Note:

What's the overall impression?
What lines feel strong or weak?
Where does attention wander?
What's memorable after one pass?

不带分析地聆听，记录：

整体印象如何？
哪些句子感觉有力或薄弱？
注意力在哪里分散？
听过一遍后记住了什么？

Step 2: Identify Primary State

步骤2：确定主要状态

Which of the six states best describes the main problem?

LY1: Feels generic/predictable → Cliche Trap
LY2: Doesn't fit the melody → Meter Mismatch
LY3: Says nothing deep → Shallow Surface
LY4: Tone feels wrong → Register Mismatch
LY5: Sounds awkward → Sound Neglect
LY6: No structure/hook → Structural Chaos

六种状态中哪一种最能描述核心问题？

LY1：感觉平庸/可预测 → 陈词滥调陷阱
LY2：与旋律不契合 → 韵律不匹配
LY3：无深层含义 → 表意肤浅
LY4：语气违和 → 语体不匹配
LY5：发音生硬 → 发音忽视
LY6：无结构/钩子 → 结构混乱

Step 3: Run Diagnostics

步骤3：执行诊断

For the primary state:

Answer the key questions
Run relevant scripts (
```
cliche-check.ts
```
,
```
meter-check.ts
```
, etc.)
Complete the diagnostic checklist

针对主要状态：

回答核心问题
运行相关脚本（
```
cliche-check.ts
```
、
```
meter-check.ts
```
等）
完成诊断检查清单

Step 4: Check Secondary States

步骤4：检查次要状态

Often multiple states co-occur:

Cliche (LY1) often comes with Shallow (LY3)
Meter (LY2) often comes with Sound (LY5)
Register (LY4) often comes with Structure (LY6)

通常多个状态同时存在：

陈词滥调（LY1）常伴随表意肤浅（LY3）
韵律不匹配（LY2）常伴随发音忽视（LY5）
语体不匹配（LY4）常伴随结构混乱（LY6）

Step 5: Prioritized Intervention

步骤5：优先干预

Fix the most damaging issue first:

Meter (LY2) before Sound (LY5)—if it's unsingable, sound texture won't help
Cliche (LY1) before Shallow (LY3)—fresh words enable depth
Structure (LY6) before Register (LY4)—need sections before matching tone

先修复最具破坏性的问题：

先解决韵律（LY2）再处理发音（LY5）——如果歌词不适合演唱，发音质感毫无意义
先解决陈词滥调（LY1）再处理表意肤浅（LY3）——新鲜用词才能承载深度
先解决结构（LY6）再处理语体（LY4）——先有清晰段落再匹配语气

Anti-Patterns

反模式

1. Over-Revision

1. 过度修改

Pattern: Fixing one problem introduces new problems. Every change cascades.

Signs:

Draft 5 is worse than draft 2
"Fixing" rhyme breaks meaning
Original spark is lost

Why It Fails: Lyrics are an interconnected system. Point fixes ignore system effects.

Fix: Make one change, then re-read holistically. Don't chain fixes without checking the whole.

模式： 修复一个问题却引入新问题。每一次修改都引发连锁反应。

迹象：

第5版比第2版更差
“修复”押韵却破坏了表意
最初的灵感消失

失败原因： 歌词是一个相互关联的系统。单点修复忽略了系统影响。

修正方案： 只做一处修改，然后整体重读。不要在不检查整体的情况下连续修改。

2. False Sophistication

2. 伪高深

Pattern: Chasing complexity for its own sake. Obscurity mistaken for depth.

Signs:

Lyrics need explanation
Dense metaphor without clear thread
Audience confusion, not intrigue

Why It Fails: Obscure isn't deep. Listeners need access point.

Fix: Every obscure element needs a clear anchor. Complexity serves clarity, not replaces it.

模式： 为复杂而复杂。将晦涩误认为深度。

迹象：

歌词需要额外解释
密集隐喻却无清晰主线
听众感到困惑而非好奇

失败原因： 晦涩不等于深度。听众需要一个切入点。

修正方案： 每个晦涩元素都需要一个清晰的锚点。复杂是为了服务清晰，而非替代清晰。

3. Genre Cosplay

3. 流派模仿秀

Pattern: Adopting genre vocabulary without understanding function.

Signs:

Country lyrics with trucks but no storytelling
Hip-hop wordplay without flow
Indie obscurity without emotional core

Why It Fails: Genre vocabulary is shortcut to function. Without function, it's costume.

Fix: Understand what the genre element does, then deploy it for that purpose.

模式： 照搬流派词汇却不理解其功能。

迹象：

乡村歌词有卡车却无叙事
嘻哈有文字游戏却无流畅感
独立音乐晦涩却无情感内核

失败原因： 流派词汇是功能的捷径。没有功能，词汇只是装饰。

修正方案： 先理解流派元素的作用，再为该目的使用它。

4. First-Draft Attachment

4. 初稿执念

Pattern: Refusing to revise lines that came easily.

Signs:

"But that's what I wrote first"
Defending weak lines emotionally
Unable to kill darlings

Why It Fails: First drafts capture intuition but include noise. Signal needs extraction.

Fix: Save first draft separately. Revise as if someone else wrote it.

模式： 拒绝修改轻易写出的句子。

迹象：

“但这是我最初写下的内容”
情绪化地维护薄弱句子
无法割舍“得意之笔”

失败原因： 初稿捕捉直觉但包含噪音。需要提取有效信号。

修正方案： 单独保存初稿。像修改别人的作品一样进行修订。

Integration

集成指南

Inbound (when to use this skill)

适用场景（何时使用本技能）

After generating AI lyrics (Suno, etc.)
When human-written lyrics feel "off"
Before finalizing any lyrics for production

生成AI歌词后（如Suno等工具）
人类创作的歌词感觉“不对劲”时
歌词定稿用于制作前

Outbound (what to do after)

后续操作（使用后该做什么）

LY1 (Cliche) → Apply Cliche Transcendence framework
LY2 (Meter) → Use meter-reference.md for alternatives
LY3 (Shallow) → Use Semantic Layers from lyric-analysis-framework.md
LY4 (Register) → Check lyricist-dna-framework.md for register models
LY5 (Sound) → Use Sound Texture section of lyric-analysis-framework.md
LY6 (Structure) → Use Structural Architecture section

LY1（陈词滥调） → 应用陈词滥调突破框架
LY2（韵律） → 参考meter-reference.md寻找替代方案
LY3（表意肤浅） → 使用lyric-analysis-framework.md中的语义分层方法
LY4（语体） → 查看lyricist-dna-framework.md中的语体模型
LY5（发音） → 使用lyric-analysis-framework.md中的发音质感部分
LY6（结构） → 使用结构架构部分

Complementary

互补技能

prose-style - for sentence-level craft (overlaps with sound/clarity)
cliche-transcendence - for escaping default patterns
naming - for sound symbolism and word selection

prose-style - 用于句子层面的打磨（与发音/清晰度有重叠）
cliche-transcendence - 用于突破默认模式
naming - 用于发音象征和用词选择

AI Lyric Generation Failure Modes

AI歌词生成的失败模式

AI-generated lyrics (Suno, etc.) fail in predictable ways due to how LLMs work:

AI生成的歌词（如Suno等工具）因LLM的工作机制存在可预测的问题：

The Degradation Pattern

质量退化模式

Lines 1-2: Model has freedom, may find something interesting
Lines 3+: Rhyme constraint activates, model optimizes for completion over originality
By verse 2: Fully in statistical default mode—stock phrases, lazy rhyme clusters

Why this happens:

Two forces compound:

Statistical median pull: LLMs generate the most probable next token. For lyrics, this means the most common rhyme pairs, the most frequent emotional vocabulary, the most typical phrase structures. Probability optimization produces mediocrity.
Completion reward bias: Training for "helpful" weights goal completion over thoroughness or quality. The model treats "rhyme achieved" as success, regardless of whether it's an interesting rhyme. Structure complete > structure meaningful.

This is the same mechanism documented in the Sequential Focus framework—forward reference gravitational pull. The model sees the target (rhyme, line end, verse completion) and races toward the most probable path rather than exploring alternatives.

第1-2行： 模型有自由度，可能产生有趣内容
第3行及以后： 押韵约束生效，模型优先完成任务而非追求原创
到第二段主歌： 完全进入统计默认模式——套话、偷懒式押韵集群

原因：

两种力量共同作用：

统计中位数引力： LLM生成最可能的下一个token。对于歌词，这意味着最常见的押韵组合、最频繁的情感词汇、最典型的短语结构。概率优化催生平庸。
完成奖励偏差： “有用性”训练更看重任务完成度而非深度或质量。模型将“完成押韵”视为成功，不管押韵是否有趣。结构完整 > 结构有意义。

这与Sequential Focus框架中记录的机制相同——向前参考的引力。模型看到目标（押韵、行尾、主歌完成）就会直奔最可能的路径，而非探索替代方案。

Observable Symptoms

可观测症状

Opening verse shows craft, later verses collapse into cliche
Rhyme words cluster in predictable groups (fire/desire/tired appears together)
Stock phrases concentrate in verses 2+
Bridge is often weakest section (furthest from opening, most constrained)

开头主歌有打磨痕迹，后续主歌崩塌为陈词滥调
押韵词集中在可预测的集群中（fire/desire/tired同时出现）
套话集中在第二段主歌及以后
桥段通常是最弱的部分（离开头最远，约束最强）

Countermeasures

应对措施

Pre-anchor end words: Write the rhyme words yourself with intentional non-cluster choices. Force the model to land on words without lazy partners. If you avoid "fire," you break the fire/desire/tired cluster.

Seed with concrete imagery: Opening lines set the vocabulary space. Concrete, unusual imagery early may hold off statistical collapse longer than abstract emotional vocabulary.

Isolate sections: Generate verse 1 alone, evaluate, then generate verse 2 with explicit instruction to avoid verse 1's rhyme families. Treat each section as a separate generation task.

Blacklist clusters: Explicitly forbid the lazy rhyme clusters in prompts:

"Do not use: fire, desire, tired, higher, liar"
"Do not use: inside, died, alive, hide, tried"
"Do not use: heart, apart, start, part"

Provide slant rhyme alternatives: Give the model escape routes. "Instead of perfect rhymes, use slant rhymes like home/come, love/move."

预先锚定结尾词： 自己选择押韵词，刻意避开集群。迫使模型落在无偷懒搭档的词上。如果避开“fire”，就能打破fire/desire/tired集群。

用具体意象引导： 开头句设定词汇空间。早期使用具体、不寻常的意象可能比抽象情感词汇更能延缓统计崩塌。

拆分段落生成： 单独生成第一段主歌，评估后再生成第二段主歌，明确要求避开第一段的押韵家族。将每个段落视为独立生成任务。

拉黑押韵集群： 在提示词中明确禁止偷懒式押韵集群：

“禁止使用：fire, desire, tired, higher, liar”
“禁止使用：inside, died, alive, hide, tried”
“禁止使用：heart, apart, start, part”

提供斜韵替代方案： 给模型提供逃避路径。“不要用完美押韵，使用斜韵如home/come, love/move。”

The Isolation Principle

隔离原则

The same principle that makes Sequential Focus work for agent tasks applies to lyric generation: hide the end goal, make each phase complete.

Don't prompt "write a complete song about X." Instead:

Generate hook/title concepts (no full lyrics)
Generate verse 1 alone (no chorus yet)
Generate chorus treating verse 1 as context
Generate verse 2 with explicit differentiation requirement
Generate bridge with perspective shift requirement

Each step is the complete mission. The model can't race to "song complete" if it doesn't know a song is the goal.

适用于Agent任务的Sequential Focus原则同样适用于歌词生成：隐藏最终目标，让每个阶段都成为独立任务。

不要提示“写一首关于X的完整歌曲”。而是：

生成钩子/标题概念（不写完整歌词）
单独生成第一段主歌（不提及副歌）
以第一段主歌为上下文生成副歌
生成第二段主歌，明确要求与第一段区分
生成桥段，明确要求转换视角

每个步骤都是独立完整的任务。如果模型不知道最终目标是写歌，就不会直奔“完成歌曲”的路径。

What This Skill Does NOT Do

本技能不做以下事项

Does not write lyrics for you (diagnostic only)
Does not evaluate musical elements (use Musical DNA for that)
Does not replace human judgment about artistic intent
Does not guarantee commercial success (craft ≠ commerce)

不为你写歌词（仅做诊断）
不评估音乐元素（使用Musical DNA）
不替代人类对艺术意图的判断
不保证商业成功（打磨≠商业）

Scripts (Pre-Filters, Not Diagnostics)

脚本（预筛选工具，非诊断工具）

These scripts surface candidates for LLM/human interpretation. They do NOT make diagnostic judgments. Use their output as input to your analysis, not as the analysis itself.

Why pre-filters?

Word lists can't catch stock phrases ("screaming inside")
Phoneme matching can't judge rhyme quality (gerund-gerund is lazy)
Syllable counting works, but variance meaning requires interpretation

这些脚本提供候选内容供LLM/人类解读。它们不做诊断判断。将其输出作为分析的输入，而非分析本身。

为何使用预筛选工具？

词表无法识别套话（如“内心嘶吼”）
音素匹配无法判断押韵质量（动名词-动名词押韵是偷懒）
音节计数有效，但差异的含义需要人工解读

cliche-check.ts

cliche-check.ts

Surfaces Tier 1/2 terms from

lyric-dominance-rules.json

. You must evaluate in context.

Usage:

bun run scripts/cliche-check.ts "your lyrics here"

Limitations: Misses stock phrases, rhyme-driven word choice, semantic cliche.

从

lyric-dominance-rules.json

中提取一级/二级术语。你必须结合上下文评估。

使用方法：

bun run scripts/cliche-check.ts "你的歌词内容"

局限性： 无法识别套话、为押韵选词、语义陈词滥调。

meter-check.ts

meter-check.ts

Reports syllable counts, stress patterns, and variance. You must judge if variance is intentional.

Usage:

bun run scripts/meter-check.ts "your lyrics here"

Useful for: Identifying lines that need review, spotting high variance.

报告音节数、重音模式和差异。你必须判断差异是否是刻意为之。

使用方法：

bun run scripts/meter-check.ts "你的歌词内容"

用途： 识别需要检查的句子，发现高差异部分。

rhyme-check.ts

rhyme-check.ts

Maps rhyme scheme and identifies rhyme types. You must evaluate rhyme quality.

Usage:

bun run scripts/rhyme-check.ts "your lyrics here"

Limitations: Marks gerund-gerund (-ing/-ing) as rhymes. Marks same-semantic-field pairs as rhymes. Doesn't detect rhyme-forced word choice.

映射押韵格式并识别押韵类型。你必须评估押韵质量。

使用方法：

bun run scripts/rhyme-check.ts "你的歌词内容"

局限性： 将动名词-动名词（-ing/-ing）标记为押韵。将同语义场词汇标记为押韵。无法检测为押韵而强行选词的情况。

Quick Reference

快速参考

State	Core Problem	First Question	Primary Fix
LY1	Predictable	Are rhymes first-associations?	Replace cliche terms
LY2	Unsingable	Do stresses match beats?	Count syllables, swap words
LY3	Empty	Is there subtext?	Add concrete imagery
LY4	Mismatched	Does tone match music?	Align vocabulary to intensity
LY5	Harsh	How do words feel in mouth?	Map sounds to emotions
LY6	Aimless	Where's the hook?	Define and place hook

状态	核心问题	首要问题	核心修正方案
LY1	可预测性强	押韵是否是第一联想词？	替换陈词滥调词汇
LY2	不适合演唱	重音是否匹配节拍？	计数音节，替换词汇
LY3	表意空洞	是否有潜台词？	添加具体意象
LY4	语气违和	语气是否匹配音乐？	调整词汇强度以匹配音乐
LY5	发音刺耳	词语的发音感受如何？	将发音与情感匹配
LY6	无方向感	钩子在哪里？	定义并放置钩子