Prompt Grammar and Section Tags
Treat a prompt as clauses, fixing the skeleton, then constraints, then exclusions, in that order.
Related
Commas are not grammar
A prompt like "warm, vintage, strings, slow, cinematic, no drums" leaves the model guessing by word frequency. Prompts have structure: section tags say what shape the song takes, control words say what must be exact, exclusions say what to leave out. Put them in the wrong order and later words overwrite earlier ones.
Tags and control words
| Element | Example | Placement | Job | Common misuse |
|---|---|---|---|---|
| Section tag | [Intro] [Verse] [Chorus] | Between lyric lines | Marks form and entry points | Tag language differs from lyric language |
| Tempo | 92 BPM | Opening | Locks the pulse | Buried at the end of a long description |
| Key | D minor | After tempo | Narrows harmonic colour | Conflicts with the reference key |
| Ensemble | String quartet, no drums | Middle | Sets timbre and density | Stacking synonyms for one instrument |
| Exclusion | no drums, no vocals | Last | Removes unwanted elements | Contradicts a positive term |
| Length | 30 seconds | Before form | Controls trimming | Does not match the section count |
The order of constraints
Genre and ensemble first, then tempo and key, then sections and length, then mood and exclusions. Earlier words act as skeleton, later words as trim. Put "no electronic drums" after "electronic drums" and the model may still see the latter first. Change one constraint per section so you learn which line did the work.
A reusable skeleton
[Style] 90 BPM, A minor, [ensemble], [mood]. [Intro] 4 bars. [Verse] ... [Chorus] ... No [unwanted].
Brackets are human-facing boundaries, not a syntax every model understands. If they are ignored, use plain lines or blank-line paragraphs; do not let the notation become noise. Keep lyrics and tags on separate lines so structure reads as structure rather than lyric content.
When the grammar fails
Structure collapses when tags are unrecognised, tags and lyrics use different languages, constraints contradict, or the prompt exceeds the effective length. Cut half the modifiers, keep tempo, ensemble and sections, then add items back one at a time. Grammar reduces ambiguity; it is not a spell that improves with length. If a new model still ignores tags, express form through line order and repeated choruses instead of fighting the notation.
Listen in the library
All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).
Along the same line
Sources
- Suno 与 Udio 官方提示指南中关于段落标记与控制词的用法。
- 流行歌曲写作教材中关于段落划分与句法的公开章节。
- 编曲实践中关于速度、调性与配器的记写惯例。