This is the only hard gate in this library: if the five items are not all filled in, you are not allowed to write the first line of lyrics.
★ Why this gate exists
The problem with AI-written lyrics is not "poor word choice," it is that "no one has made a choice."
It will write four sections of lyrics that are grammatically correct, neatly rhymed, and rich in imagery, and these four sections could belong to any song. That is exactly the reason it sounds like AI.
These five questions are where you are forced to make choices. The latter four items are the same kind of questions you must answer before writing music or writing drama - first choose what to say, to whom, and at which moment; the first item is unique to lyrics, becauselyrics have types, music does not。
Which sections to read according to the task
| Task | Read |
|---|---|
| Starting a new song | §1 → ★ first answer §1.1 to determine the type → answer §2–§5 one by one |
| Everything written sounds like it is telling a story | ★ §1.1、§1.2（The default is not narrative type） |
| Cannot answer | ★ §6（Being unable to answer is itself information） |
| Cannot tell whether you really answered or just filled it in | ★ §7 |
| After the lyrics are written, they feel generic | Come back and answer again |
Boundaries
| Not covered here | Where it belongs |
|---|---|
| How to find imagery, practice methods for object writing | lw-imagery |
| Thetechniques | lw-narrativeof perspective, person, and time anchor (★ this skill only asks "at which moment"; it handles "how to write that moment") |
| Section structure | lw-structure |
| Sociological/literary analysis of themes | Not done |
1. The five questions
intent:
plot_type: # ★ 先答。这首歌是哪一类：attitudinal / situational / narrative / expository
one_thing: # 这首歌只说一件事
to_whom: # 对谁说
at_what_moment: # 在哪个时刻说
anchor_object: # 一个贯穿全曲的、能拍照的物件
★ The five items are not parallel to one another：
plot_type defines the framework,one_thing defines the content,to_whom defines the tone,at_what_moment defines the tense and amount of information,
anchor_object defines the vocabulary. The latter four items are all answered within the framework set by the first item, and the latter three items in turn revise one_thing。
1.1 ★ plot_type —— first determine which category this song is
Davis lists "choosing the plot type" as thefirst stepin constructing a song, before deciding the emotion, arranging the elements, and drawing the conclusion (p0025). The other four questions are all answered within one type; if the type is not set, the answers to the four questions will each drift.
Closed set, choose one of four：
plot_type |
What it is doing | Whether there is a timeline | Example |
|---|---|---|---|
| attitudinal | The singer expresses an attitude or emotion toward someone or something | No | ★ Davis：Most pop songs belong to this category（p0025） |
| situational | Gives the attitude a dramatic frame; the singer isreacting to a specific situation | One scene, no progression | Farewell; a ex-convict riding the bus home (p0025) |
| narrative | A real plot, linear, with a beginning and an end | Yes | Recalling personal events; talking about people one knows; a "camera-eye view" without person marking (p0025–0026) |
| expository argumentative type | Seize on one idea and say it over and over | None | In the Chinese tradition it stands as a category of its own. ★ Wu Songjin: songs are not good at arguing a point,you have to turn philosophy into wit, then combine it with lyricism（p0049） |
★★ By default it is not narrative type. Davis says most pop songs are attitude type or situation type (p0025, p0061), Wu Songjin says lyricsare more lyrical and less narrative, and even when there is narrative it is passed over in a stroke (p0034). The two traditions independently arrive at the same conclusion.
Corpus test(stratified sample of the top 100 singers on the Chinese-language singer chart, each song typed according to the three questions above): Mandarin non-rap attitude type 64%, situation type 24%, argumentative type 8%, narrative type 5%(n=487); rap 67/22/3/8 (n=100); Cantonese attitude type 69% (n=36, there are only this many Cantonese songs in the corpus, the other cells are too small). The judgments of the two traditions hold on the Chinese-language corpus, and narrative type is one song in twenty. It does not move by era: in the 2000s, 2010s, and 2020s cells, attitude type is 60% to 67%, narrative type 3% to 6% (n=73/154/245). Recheck: 59 songs were sampled and blind-judged by another model, type judgments agreed 75%, and for those reporting high confidence 87%; the disagreements were almost all between attitude type and situation type, while the shares of narrative type and argumentative type were stable.
Why words without a chosen type inevitably slide toward narrative: narrative is the only one of the four types that can advance just by "continuing to tell", while the other three all rely ontaking another lookto advance. If you do not know which type you are writing, every section will want to write something new, and the new can only be new events, so every song becomes a story.
A song written by repeatedly circling one marker (one action, one object, one catchphrase) is not a fifth type. That is attitude type or situation type handing the whole structure over to
anchor_object(§5.3), with the object doing what a section should do.
1.2 ★ A chain of downstream choices it determines
This table is the thing most easily missed after the four questions are answered:Once the type is set, person, voice, and repetition budget are set along with it。
| Person | Voice | Repetition budget | Main skill to go to | |
|---|---|---|---|---|
| Attitude type | First and second person, emphasizing I and you (p0061) | Talking to oneself, or speaking to someone | The chorus repeats word for word;the verse can also change only the angle without changing the matter | lw-imagery: sensory words are arranged from far to near (p0027). ★ Four out of ten attitude-type songs have no object, and the anchor is one sentence (start of §5) |
| Situation type | First and second person (p0061) | ★ It must be locked down whether one is speaking or thinking, if not locked down the words blur (p0069) | The chorus repeats word for word; the verse relies ontaking another lookto advance | anchor_object to carry the structure (§5.3);lw-narrative leaving blank space |
| Narrative type | Third person is characteristically used for story songs (p0061); first-person recollection is also possible | Told to an unseen audience, in essence it is "thinking" (p0069) | ★ The chorus still repeats word for word; only the verse advances through new events | lw-narrative all;lw-structure |
| argumentative | first/second person | speaking to someone | chorus repeats word for word; verse changes examples but not the argument | the risk is a preaching tone. go back to §5 to find an object, turn the argument into something interesting |
In the corpus, 86% of situational type have a through-line object, while only 60% of attitudinal type do (n=115/313); the principle that an object carries the structure is the norm for situational type; first person in narrative type is only 59%, lower than the other three types, but third person across the whole sample is only 2%, so "story songs use third person" is a minority practice in Chinese.
The person column is lw-narrative the basis for choosing cells in the WTS table in §1.9, the table is over there, which cell to choose is here.
Look carefully at the repetition budget column: the choruses of all four types repeat word for word, this pointdoes not change with type;
type only determineswhat the verse uses to advance. Narrative type is the only one that advances the verse through new events, the other three types all rely on taking another look.
In the tests, the two narrative commercial songs were actually the most repetitive, because they handed the entire chorus over to repetition, and only then did the verse have room to tell the story.
So the chain of problems for lyrics with no chosen type is: don't know the budget → don't dare repeat the chorus either → every section needs new content → can only rely on new events → the lyrics are long and all story.
On how to spend the budget, see lw-structure §6.0。
1.3 How to judge, and how to know you judged correctly
Ask three questions in order：
- Does this songhave a timeline(has a beginning and an end, events are advancing)? Yes → narrative type
- No timeline, butthere is a specific scene, and the singer is reacting to it? → situational type
- Neither timeline nor scene, justone attitude or one argument repeated？→ attitudinal typeorargumentative type
The failure mode of situational type(Davis p0117): listeners say "like a theater song", meaning this song depends on a script, and the script is only in the author's head, it isn't written in the lyrics. Listeners don't know under what circumstances this is happening: on stage? at a party? in daily life? Criterion: whether the situation can be read from the lyrics themselves. If it can't be read out, go back to §4 and write the moment in.
The criterion for judging "really answered": can statewhy the second verse is not a restatement of the first verse, and the answer must match the type. Narrative type answers "something new happened"; the other three types answer "the same thing from a different angle, with added weight". If the answer is "something new happened" but the type filled in is attitudinal type, one of the two was filled in.
2. one_thing —— this song says only one thing
2.1 Specification
- one sentence
- two sentences not allowed
- ★ "both... and...", "on one hand... on the other hand..." not allowed
Davis lists "conciseness" as the first of the ten major writing principles:stick to one idea, eliminate side plots, a carefully crafted lyric plot,should be able to be summarized in one short sentence(p0086). The counterexample she gives is a student writing about a high school girl secretly in love with a classmate, writing from the first day of second grade to the second week of eighth grade, treating "complexity of plot" as "development of plot," all irrelevant details; the fix is not to cut details, but tokeep only the moment the boy ignores her, letting one line of the chorus carry all the remaining emotion (p0086). This criterion can be used in reverse as well:if you can't sum it up in one sentence, it means you're writing about a process, not this thing itself.
2.2 The failure mode it catches
A song wants to say two things → it says neither of them fully.
This is especially common in AI-written lyrics, because the model writes in every relevant facet of a theme: love becomes "longing + regret + letting go + blessing," one per verse,after listening you don't know what it's saying。
2.3 ★ How to judge whether it's really one thing
Say it out loud.
| Symptom | Judgment |
|---|---|
| Sounds like a human sentence when said out loud | ✅ |
| Sounds like a thematic word ("love," "growth," "loneliness") | ❌ That's a subject, not a premise |
| Has an "and" in the middle, or a comma that can split it into two halves | ❌ Two things |
| After saying it, you want to add "and also…" | ❌ |
Counterexample → positive example：
| ❌ | ✅ |
|---|---|
| About heartbreak and growth | The day she moved out I realized the balcony was mine |
| Expressing longing for one's hometown | My mom only asks if I've eaten when she calls, and I always say I have |
| The beauty and regret of youth | There's someone in the graduation photo whose name I still can't remember |
★ Note what the positive examples have in common:They all already carry an image, andthey all have a touch of the unseemly。
3. to_whom —— To whom it's said
3.1 Specification
A specific target. ★ "Lyricism" is not a target. "The listener" is not a target either.
Davis calls this target singee: the one person, place, or object uniquely addressed as "you" among second-person words,
present, absent, or abstract all count,but once chosen, you have to acknowledge it all the way through, and no other "you" is allowed to mix in and confuse the listener（p0064）。
to_whom What gets filled in here is the name of this singee; if you can't fill in a specific name, it means this item hasn't actually been chosen yet.
3.2 The failure mode it catches
A lyric with no target automatically slides toward "correct words addressed to everyone"—— that kind of talk is inevitably abstract, inevitably safe, inevitably AI-like.
Davis nails this down from the negative side: a preachy message addressed to a generic "you"inevitably feels preachy(p0065), what the listener wants isto eavesdrop ona sentence spoken to some specific person, not to be lectured face-to-face by a crowd. So even if the song really does want to make a point (the argumentative type is the same), changing "you" to "everyone" won't solve the preachy feeling;the method that really works is to narrow the target down to one specific person and let the listener eavesdrop（p0066）。
3.3 Possible answers
| Type | Example |
|---|---|
| A person | An ex three years after the breakup; my sixteen-year-old self; a grandfather I've never met |
| A person who is not present | ★ Best to use—becauseyou can say what you wouldn't dare say to their face |
| An object | That car that was sold; a house that has already been demolished |
| A specific group of people | The cohort that graduated the same year I did (not "young people"） |
| Yourself | —but you have to make clear it'swhich version of yourself |
3.4 ★ How it changes the words
The same thing, with a different addressee, becomes completely different words:
| Addressee | Tone |
|---|---|
| Saying it to an ex | There are scores to settle, and there will be specific details |
| Saying it to your current partner | You have to be careful, and there will be omissions |
| Saying it to yourself | You don't have to be reasonable |
| Saying it to someone who will never hear it | ★ The freest, you can say it outright |
Self-check action: Imagine this lyric as a message that has been sent.Sent to whom? If you can't answer, the lyric still has no addressee.
4. at_what_moment — at which moment it is said
4.1 Specification
★ A point in time, not a span of time.
| ❌ Time span | ✅ Point in time |
|---|---|
| After the breakup | The day of the move, the last trip downstairs |
| That summer | Three o'clock on the afternoon I found out the results |
| After growing up | The moment I paid rent by myself for the first time |
4.2 The failure mode it catches
A lyric with no moment is a "summary"—it is summarizing a stretch of experience, not speaking from a particular position. A summary is inherently abstract, because it has to cover an entire period of time.
Davis's criterion:Unless you tell the listener that the time frame has changed, a song by default has only a single moment's feeling(p0076),
time cannot jump far forward between sections without that jump being written out.
The counterexample she gives is a children's song that begins "Mommy brought her home today," but by the last verse the older brother is pushing his little sister in a stroller around the living room,
and listeners get stuck on "how can a newborn ride in a stroller?" The problem is thatthe writer forgot to make clear across the four verses that three years had passed(p0076).
An example that does it right is "Cat's in the Cradle," where each verse begins by marking a line like "the day my son learned to walk,"
"the day my son just turned ten,"the advance of the years is written out explicitly, so it doesn't sound like a summary（p0076）。
at_what_moment Answering with a point in time essentially means first locking the default to "a single instant";
to cross time, you have to write the jump explicitly like this, otherwise the lyric will slide on its own into a boundaryless retrospective.
4.3 ★ It determines how much information there is
A specific moment limits what you can know.
Example: the intent is "we will break up," and the moment is "some night while we are still together"— thenthe lyric cannot contain "we broke up", only what could be observed that night (he took longer than usual washing the dishes, replied more slowly on WeChat, set the air conditioner to the temperature he likes).
★★ This constraint is precisely the mechanism that makes the lyrics better. It forces you to usesymptomsinstead ofconclusions, and symptoms are concrete nouns.
→ This one connects directly to lw-narrative 's "leaving blank space" and lw-imagery 's "concrete nouns".
5. anchor_object — an object you can photograph
★ The anchor can also be a single line. In the corpus, among the 313 Mandarin attitude-type songs, forty percent have no object running throughout (66 songs don't even have an imagery system, and 59 have only atmosphere with no object). This group's approach is very consistent: the chorus isa line spoken directly to "you"(shouting out a breakup, begging you not to leave, demanding to know if you dare, paired rhetorical questions, slogans, short self-blaming lines), and the verse sayswhat state I'm in right now, giving no imagery; 41% of the hook lines are monologues, 14% are questions; colloquial language accounts for 82% (only 54% across all attitude-type songs); 52% of final choruses stay exactly the same; the second verse often repeats word for word or changes only a word or two. To write this kind of song,
intent.anchor_objectfill in "none" foranchor_phrase, and the criteria for objects discussed below in this section don't apply; for the repetition budget followlw-structure§6.0. Use the criterion in reverse:if you've written an object and imagery, but the brief calls for "saying it directly", then you've written it as the situational type.
5.1 Specifications
★ Can be photographed.
This is a very harsh criterion; it eliminates all at once: time, distant places, dreams, freedom, youth, longing, tenderness, courage, stars and seas...
| ❌ | ✅ |
|---|---|
| memory | a movie ticket pressed under glass |
| distance | that address still not deleted from the phone |
| time | that sweatshirt he left behind, washed until it pilled |
| longing | that extra half bowl of rice always cooked |
5.2 The failure mode it catches
Piling up abstract nouns— the most direct source of AI flavor.
Davis lists this as the "concreteness" of the ten principles: listeners want to see, hear, touch, smell, feel,
abstract words like kindness, sorrow, memorieshave neither human warmth nor shape, and leave not a trace in the listener's memory(p0095).
Her approach is to replace "shadows" with "physical objects":memories of you replaced with that velvet valentine，
jewelry you wore replaced with gold bangle bracelets, and she calls this kind of photographable word Kodak words（p0095）。
anchor_object 's "photographable" criterion is another way of saying Kodak words:
if a line still holds when replaced with an abstract noun, that means it isn't an object, it's an emotion with a label stuck on it.
5.3 ★ Why it must "run through the whole song"
If an object appears only once, it's a detail; if it appears three times and means something different each time, it's structure.
Example (that sweatshirt)：
| Section | What it is |
|---|---|
| Verse 1 | a piece of clothing left behind |
| Chorus | something that still holds a scent |
| Verse 2 | should have been returned but never was |
| Outro | washed, the scent is gone |
★ The object doesn't change, but its meaning evolves as the song progresses — that's what "running through" means. This is also the most economical structural device in lyrics:No explanation is needed; you just need to let it appear once more.
5.4 A clue to hand to composition and arrangement
★ anchor_object It can often also suggest the arrangement's hook—
A recurring object paired with a recurring sound event.
How to make the arrangement's hook is a matter for composition and arrangement; what this library hands over is intent.anchor_object, and in which lines it appears.
6. ★ What to do if you can't answer
Not being able to answer is not "haven't thought it through yet"; it is information.
| Which item you can't answer | Explanation | What to do |
|---|---|---|
| one_thing | You haven't yet figured out what you want to write | ★ Don't start writing. First go experience/recall/observe |
| to_whom | You are writing "some kind of emotion in a general sense" | Pick a specific person, even if fictional |
| at_what_moment | You are writing a period of time rather than a location | Pick from that period of timethe most painful or most awkward minute |
| anchor_object | ★ Most common | See §6.1 |
6.1 When you can't find an object
Go back to that moment and ask three questions：
- What were you holding in your hand at that time?
- What in that room was something others wouldn't notice?
- Is there anything that is still with you now?
★ The 3rd one works best—something that is still with you naturally has a "running through" quality.
→ For a more systematic method, see lw-imagery object writing.
7. ★ How to tell "really answered" from "filled in to pass the check"
This is where this skill most needs self-discipline. Four criteria:
| # | Criterion |
|---|---|
| 1 | one_thing When spoken aloud, it is a human sentence, not a phrase |
| 2 | to_whom It is a person whose face you can imagine |
| 3 | at_what_moment Precise to the minute level, not "some day" |
| 4 | ★ anchor_object It can be photographed, and you can state its color, size, and whether it is new or old |
★★ The 4th one is the demon-revealing mirror: A filled-in object can't have its color or newness stated, because it was reverse-engineered from the need that "there should be an object," not taken out of memory。
8. Correspondence with LYR-SPEC fields
| Field | This skill |
|---|---|
intent.plot_type |
★ §1.1, choose one of four from a closed set |
intent.one_thing |
§2 |
intent.to_whom |
§3 |
intent.at_what_moment |
§4 |
intent.anchor_object |
§5 |
imagery.forbidden |
★ Those words excluded in §5.1,write into the blacklist |
imagery.concrete_nouns |
anchor_object is one of them, but not the only one |
narrative.tense_anchor |
★ Determined by at_what_moment determined |
narrative.pov |
★ Determined by plot_type to determine the broad category (§1.2), then by to_whom to determine the specific object |
asymmetry.repetition_budget |
★ Determined by plot_type to determine (§1.2); for how to spend it, see lw-structure §6.0 |
★ LYR-LINT item 1 checks these five，plot_type recognizes only four values;anchor_object whether it "can be photographed" depends on human judgment: only those whose color, size, and newness can be stated count.
9. Self-check before passing the gate
- Five itemsAll filled in
- ★
plot_typeYesDetermined by the three questions in §1.3, it is not the default narrative type; you can explain why the second paragraph is not a restatement of the first -
one_thingread aloud, it sounds like something a person would say, with no "and" and no comma splitting it in two -
to_whomisan object whose face you can imagine, not "listeners"/"lyrical expression" -
at_what_momentisa point in timenot a time period -
anchor_objectcan be photographed, and you can state itscolor, size, new or old - ★ Have thought about what this object meansin each section respectively（§5.3）
- ★
at_what_momentlimited what can appear in the lyrics (§4.3)——did not write things that were not yet known at that moment - put the abstract words excluded by §5.1into
imagery.forbidden
Appendix: Sources and Status
plot_type's three-way distinction comes from Sheila Davis, Successful Lyric Writing(p0025–0026 definitions, p0061 person derivation, p0069 the voice of situational songs, p0117 failure modes of theatrical songs), the argumentative branch comes from the Chinese lyric-writing tradition(Wu Songjin p0034, p0049). The two traditions independently converged on the same set of types.- The framework for the other four questions was set by this library itself, based on a summary of "why lyrics written directly by AI can be recognized instantly";
but the criteria for each question can each be matched to an independent principle in Davis's book:
one_thingcorresponds to "conciseness" (p0086),to_whomcorresponds to the selection of singee and not preaching (p0064–0066),at_what_momentcorresponds to "the feeling of a single moment" and the need to state time jumps explicitly (p0076),anchor_objectcorresponds to "concreteness" and Kodak words (p0095). - ★
to_whomis a question only lyrics have: becauselyrics have a speaker, music does not。 - ⚠ The examples in §2.3, §3.3, §4.1, and §5.1 were invented by this skill, not copied from the source text.