SuperAPI AI Directory

lw-suno-interface

Coding & development

View repository

Updated 10-15 23:41

About this skill

Lyric Draft → Backend (Lyric Compile)

This skill only handles one thing: turning a finished lyric into input the backend can digest, without losing information along the way.

★ Division of labor first

Library One has a compiler mc-render-compile, which handles compilation of the entire ARR-SPEC. This skill does not duplicate it, only handles the lyric part, andfollows the rules it sets。

Who handles What
Library One mc-render-compile style_prompt, section tagssequence, exclude, seed, backend selection,honors contract
This skill ★ the format of the lyric itself: line breaks, punctuation, parentheses, phrasing within a section, alignment with tags

★ When in conflict, Library One prevails—it is the sole compiler, and this skill is one of its input sources.

Which sections to read depending on the task

Task Read
Handing lyrics to Suno §2 + §3
Phrasing is wrong when sung ★ §3
Section order is messed up §2.2
How to handle punctuation/parentheses §4
Which backends can take lyrics §5
What to check before handing off §6

Boundaries

Not handled here Handled where
Style description, section tag sequence, seed, exclude Library One mc-render-compile
Whether the lyric is well written Each L1/L2 skill in this library
How those words should be sung (breathy voice/strength/pitch correction) Library One mc-vocal-direction
Whether the generated audio sounds like AI Library One mc-ai-tell-audit
Whether the lyric itself sounds like AI lw-ai-tell-audit

1. Must-check before compiling

Must-check Why
Has LYR-LINT passed? If lint does not pass, the lyric is not finished
Are section ids aligned with ARR-SPEC.form[].id ? ★ If not aligned, the lyric will be assigned to the wrong section
Has the inverted-character check been done? For Chinese lyrics, and if a melody already exists
Does the target backend accept lyrics? §5。★ Stable Audio No vocals

2. Section tags

2.1 Mapping table

Consistent with Library One (mc-render-compile §5.1):

form[].name Tag
Intro [Intro]
Verse [Verse]
Pre-Chorus [Pre-Chorus]
Chorus [Chorus]
Bridge [Bridge]
Instrumental [Instrumental]
Outro [Outro]

★ Two traps(already written on Library One's side, repeated here because lyric writers are most likely to step on them):

  1. 大サビ maps to [Bridge] not [Chorus] — if mapped wrong, the backend will give you a high-energy chorus at the introspective position
  2. 落ちサビ (a chorus with sparse arrangement) is still [Chorus] — the sparse requirement goes into style_prompt,do not express it by changing tags

2.2 Order and count

  • ★ The number of tags in the sequence must equal form the number of sections in(Library One lint #13)
  • The lyric is split by section and corresponds one-to-one with the tags
  • If a tag is missing, the backend will decide on its own what that section is

2.3 Instrumental-only sections

[Instrumental] Belowdo not put lyrics: this is the basis for the backend assigning section positions by tag; if lyrics are stuffed in, they will be treated as sung phrases. ★ Butkeep this tag, otherwise the following sections will shift forward as a whole, one slot will be missing from the sequence, and the backend will use the next tag to fill the position this section should occupy. This is the backend's fixed assignment behavior; there are no conditions to discuss, and two sentences are complete.


3. ★ Line breaks: it is not typesetting

This is the most important section of this skill.

Line breaks within a section are read by the backend as phrase boundaries. It is not typesetting, it is aninstruction。

3.1 Consequences

What you write The backend understands as
One sentence split into two lines Two phrases, with a breath in between
Two sentences merged into one line One long phrase, with no breath in between
Blank line An even longer pause

★ Therefore, the line breaks in the lyric draft must match prosody.sections[].syllables consistently—— one element in that array is one line.

3.2 Relationship with breath points

narrative and prosody in breath_points(library one vocal.breath_points) should fall at line breaks. If inconsistent:

  • Breath point in the middle of a line → the backend will not take a breath there, and the singer will hold it in
  • No breath marked at a line break → extra pauses may occur

★ Self-check action: read the lyrics aloud line by line,take a breath at the end of each line. If you cannot read it through, the line breaks are wrong.

3.3 One character with multiple notes and melisma

The backend cannot tell which character you want to lengthen. If you want melisma, write it explicitly in the lyrics——different backends write it differently; the common way is repeating the vowel (a-a-ah) or using a hyphen. The actual behavior of each backend is subject to testing.


4. Punctuation and brackets

Element How the backend usually handles it Suggestion
Commas, periods Mostly ignored, butmay be treated as a slight pause ★ Avoid if possible——using line breaks to express pauses is more reliable
Parentheses `` ★ Often understood asharmony/backing vocals(backing vocal) Use them if you want backing vocals;do not use parentheses for annotations
Square brackets [] ★ reserved for section tags, do not use them anywhere else
Ellipses, dashes Behavior uncertain Avoid
Capitalization(English) May affect emphasis Write normally
Annotations, creation notes ★ Will be sung as lyrics ★★ Absolutely do not put them in the lyrics——put them in LYR-SPEC's note field

★★ The last one is the most common accident: If you write "(repeat here three times)" in the lyric draft, the backend willsing this sentence out loud。


5. Which backends can eat lyrics

Backend Lyrics Description
Suno ✅ Section tag + lyrics. Highest level of readiness
YuE2 ✅ ★ Can also eat ABC notation— both lyrics and melody can be given precisely
MiniMax Music ✅ ⚠ The commercial API has been closed to new users
ACE-Step ✅
Stable Audio ❌ ★ No vocals
Symbolic paths (MIDI/MusicXML) ⚠ MusicXML can carry lyrics (syllable-by-syllable alignment); MIDI can only be treated as text events

★ See Library One for details, in the mc-workflow directory's backends.yaml。

5.1 ★ YuE2's special value

It can simultaneously eatlyricsand ABC notation. This meansthe inverted-character problem can be locked down before generation— You have specified the melody, the detector has verified it, and the model sings accordingly.

Other backends can only "hint" at the melody, so the results of the inverted-character checkare not guaranteed to be delivered on them。


6. Must pass after compilation

  • LYR-LINT passed
  • section tagscount == form number of sections
  • section ids correspond one-to-one with ARR-SPEC.form[].id one-to-one correspondence
  • ★ The big chorus is mapped to [Bridge]
  • [Instrumental] sectionkept the tag but placed no lyrics
  • ★ Line breaks are consistent with prosody.syllables consistent; read line by line, each line can take a breath at the end
  • ★★ There are no comments/creative notes in the lyrics(they will be sung out loud)
  • Parentheses are used only forbacking vocals, and square brackets are used only forsection tags
  • Use as little punctuation as possible; pauses rely on line breaks
  • Target backenddoes indeed accept lyrics(not Stable Audio)
  • Chinese: inverted-character check has been done; ★ If the backend is not YuE2, be aware that the results are not guaranteed to be delivered(§5.1)

Appendix: Sources

  • Basis for the division of labor: Library One mc-render-compile(the only compiler) and backends.yaml

  • §3.3 and §4 are general experience; the actual reactions to melisma writing, punctuation and parentheses, and blank lines within a section vs single line breaks vary by backend and should be based on actual testing

  • This skill is not "Suno-specific"; it is "compilation of lyrics," aligned with the compiler architecture in Library One §2.3.