Every draft must pass through this skill before it goes out. This is the final check of this library.
Core proposition: AI-written lyrics are not "bad word choice," they are **"no one has made a choice."** The four sections it writes are grammatically correct, neatly rhymed, rich in imagery, and these four sections could belong to any song.
Which sections to read by task
| Task | Read |
|---|---|
| Just finished writing, needs acceptance | §2 the eight items → §3 |
| "Reads fine but is just generic" | ★ §2（Check item by item, don't go by feel） |
| All passed but still feels like AI | ★ §4 |
| Found problems, need to fix | §5 |
Boundaries
| Not handled here | Where it belongs |
|---|---|
| Check when the spec is finished | LYR-LINT ten items（lw-workflow §5.2） |
| How exactly to fix | Each L1/L2 skill (§5 gives routing) |
| Judging inverted tones | lw-tone-check |
★ Division of labor with LYR-LINT： lint asks "is the spec filled in," this skill asks "does this lyric read like a human wrote it." Only when both pass is it done. Lyrics that pass all ten items can still be full of AI flavor.
1. Before starting
| Required | How to ask |
|---|---|
| ① This draft's version and source | AI direct write? skill-guided? handwritten?When doing AB, this column is everything |
| ② Language | Chinese requires extra checks for inverted tones/rhyme categories |
| ③ Is there a melody | If not, item 5 cannot be checked |
2. ★ The eight-item enumeration
Check off item by item. Don't go by overall impression.
| # | AI-tell symptom | LYR-SPEC field it counters | Automatable | Who handles it |
|---|---|---|---|---|
| 1 | All nouns are abstract | imagery.concrete_nouns required,forbidden blacklist |
✅ part-of-speech and wordlist statistics | lw-imagery |
| 2 | Structure symmetrical to death | asymmetry three required items |
✅ character-count variance per section | lw-structure |
| 3 | Everything is resolved, nothing left hanging | narrative.withholding required |
⚠ semi-automatic | lw-narrative |
| 4 | No perspective or time | narrative.pov ＋ tense_anchor |
⚠ pronoun detection can assist | lw-narrative |
| 5 | Inverted tones ignored | prosody + inverted-tone detector |
✅ Fully computable | lw-tone-check |
| 6 | Rhyme categories only pick the smoothest | rhyme.che + rhyme-category distribution statistics |
✅ | lw-mandarin |
| 7 | Rhyming always picks the first character that comes to mind | rhyme.surprise_notes |
✅ rhyme-word frequency | lw-rhyme |
| 8 | Every chorus repetition is identical, and the last repetition neither changes a word nor adds a tail(being fully identical itself accounts for 27% in the corpus, not a defect) | asymmetry.chorus_variation |
✅ text diff | lw-structure §6.0 |
★ Six of the eight items are computable(items 3 and 4 are semi-automatic). This ratio is harder than comparable checks for audio, because text needs no detection.
★ This table serves three purposes at once: The skeleton of this skill, the source of the ten LYR-LINT rules, the skeleton of the evaluation table when doing an AB comparison of two drafts (
lw-workflow§7.2）。
2.1 ★ The three most fatal
Not because they are the hardest to detect, but because none of them reads as "wrong":
| # | Why fatal |
|---|---|
| 1 (all nouns abstract) | ★ The most direct AI fingerprint. "Time", "distant places", "dreams", "freedom"—each one is right, but together they are empty |
| 2 (symmetrical to death) | Four sections, four lines each, seven characters per line, every line rhymes.Technically perfect, but it sounds like a spreadsheet |
| 3 (everything gets resolved) | AI will walk the emotion all the way through: sadness → letting go → blessing.Real people writing lyrics often stop inside the pain |
2.2 Four coarse filters you can count immediately
# 抽象词命中数 / 具体名词数
# 每段字数方差
# 副歌各遍的文本 diff（★ 只看末遍有没有改动；前两遍相同不算问题）
# 韵脚字的词频排名
These four itemscount by hand—each item takes under a minute, no script needed.
3. Two numbers, must not be merged
| What it asks | How to get it | |
|---|---|---|
| LYR-LINT pass count | How much of the spec is filled in | Ten rules |
| AI-flavor flag count | How many machine fingerprints the lyrics have | Eight in this skill |
lint 全过 + 旗标少 → 好，交
lint 全过 + 旗标多 → ★ 规格本身填得像 AI（见 §4）
lint 不过 + 旗标少 → 词不错但规格没写完，补规格（为了可复现）
lint 不过 + 旗标多 → 回 lw-workflow 的 S0
4. ★ It passes everything and still feels like AI
This is the main reason this skill exists.
4.1 Four signs
| Sign | Where to check |
|---|---|
intent.one_thing Read it aloudSounds like a theme word, not human speech |
lw-song-intent §2.3 |
anchor_object Can't say what color or how old/new it is |
★ lw-song-intent The demon-revealing mirror of §7—it is derived backward, not taken from memory |
| Concrete nouns are there, butthey do not form a system(randomly piled imagery) | lw-imagery（imagery.system field) |
asymmetry Three items are filled in, butthose asymmetries have no reason |
lw-structure——★ Breaking things just to break them is as mechanical as being neat |
★ Common point：The rules are treated as conditions to satisfy, not choices to make. This is exactly the definition of AI flavor.
4.2 A harsher test
★ Remove the title and all proper nouns from these lyrics, then show them to someone who knows nothing about them. Can he say which specific thing this song is about?
Can't say it →
one_thingit did not land in the lyrics,it only lives on the spec sheet。
5. Remediation path
读着像 AI
└─ ① 逐条过 §2 的八条，数出旗标数
└─ ② 旗标 ≥ 3 → 按下表逐个改
└─ ③ 旗标 ≤ 2 但仍然像 → §4 的四个征兆
└─ ④ 还是不行 → ★ 回 lw-song-intent 重答五问
5.1 Symptom → skill
| Symptom | Check first |
|---|---|
| Empty, vague, says nothing | lw-song-intent（Mostly one_thing does not hold） |
| All nouns are concepts | lw-imagery |
| Reads like prose, not a song | lw-structure(stability and instability of sentence patterns) |
| Neat as a spreadsheet | lw-structure of asymmetry |
| Rhymes smoothly but means nothing | lw-rhyme §surprise level |
| Sung aloud, it sounds like other words | lw-tone-check |
| Don't know who is speaking | lw-narrative |
| The emotion is walked all the way through, no aftertaste | lw-narrative of withholding |
| The chorus is identical all three times, and the last time not a single word is changed | lw-structure §6.0 (★ It is normal for the first two passes to be identical; don't rewrite the chorus, change one to three characters in the last pass) |
| It doesn't sound right when sung in Cantonese | lw-cantonese（★ It's not inverted characters, it's harmonic coordination） |
6. ★ What is this checklist based on
There is only one criterion: whether this draft can be distinguished from a human-written lyric.
- is taken as "human-written" → this draft passes
- can still be distinguished from a human-written one → the gap is exactly what this draft needs to make up
For the gap you found, go through §2 of this skill item by item to find which item it is, and revise according to the skill it points to.
7. Pre-submission checklist
- item by itempassed the eight items in §2,counted the number of flags(not by impression)
- if in Chinese, raninverted-character detectionand manually confirmed the high-risk spots
- ★ did that test in §4.2:remove the title and proper nouns, and others can say what event it's about
-
asymmetryevery asymmetry incan give a reason - concrete nounsform a system, not randomly piled up
- each pass of the chorusis indeed different, and the changesare meaningful(not swapping in a synonym)
- abstract-word blacklistzero hits
- if it all passes and still feels like AI → go back to
lw-song-intentand re-answer the five questions, don't fix it at the wording level
Appendix: Sources
- skeletonis an AI-flavor symptom → the comparison table of LYR-SPEC countermeasures; the same table is also in
lw-workflow§5.1 - ⚠ This skill does not have: the complete word list of the abstract-word blacklist (only scattered examples), rhyme-category distribution statistics (the thirteen-rhyme table is in
lw-mandarinreference §1), rhyme-word frequency table - Items 2, 7, 8, and 9 only report numbers, without judging pass/fail