Edited by us
AI music prompt — why the lyrics vanish
Part labels with no words to sing, and tempo left blank. Those are the two traps of music requests.
By Mephify (operator)3 min read
What makes this hard
Music tools usually take two fields: style, which sets the sound, and lyrics. Ask AI for the prompt and one side often comes back empty: [Verse 1] [Chorus] part labels with descriptions like 'this part follows the commute-home gaze', and nothing to actually sing. Paste that in and the engine invents its own words, different on every run.
The other trap is values left blank. Leave tempo and key 'to taste' and the decision passes to the engine, which means a different song every time. An empty field isn't humility. It is a randomizer button.
What to add to the request
1. Demand both blocks: style and lyrics
Put it like this
Give me two separate blocks: one to paste into the style field, one for the lyrics field. Under each part tag write words to actually sing — don't substitute descriptions.
For instrumentals, write the opposite: 'don't create a lyrics field'. An empty lyrics box reads as 'fill me' to the engine.
2. Unknown values come back as recommendations
Genre implies a usual range for tempo and key. Ask it to propose a value with its basis: '92 BPM, the usual range for mellow city pop; say the word to go faster'. You get a decided prompt without knowing the values yourself. Same for instrumentation and dynamics: not 'stronger in the chorus' but which parts enter and which drop, as values.
3. Separate keep from avoid
- Lyric lines you provided stay verbatim — it builds around them
- No reproducing existing melodies or lyrics; no naming a specific artist as the vocal target
- Lyrics must fit the melody in syllable count and breath, not just meaning, so require even line lengths
Two blocks to paste — style and lyrics
The render section of a music prompt is one block for the tool's style field: genre, tempo, key, instruments, vocal, arrangement density, and mix joined by commas. The words to sing belong to a separate lyrics section. The two never repeat each other, and the rule to write in noun phrases applies to style only. Lyrics are what a person sings, so they are written as sentences.
The structure section shows parts and their timing, and no lyrics go there. What gets louder and what drops out in each part is written as values. 'Stronger in the chorus' is not something an engine can execute.
Keeping the same voice across an album
A vocal anchor describes a timbre; it can't pin a voice. When track-to-track consistency is the goal, the prompt writes a procedure: if the tool supports a vocal seed, a reference track, or a saved vocal profile, reuse the first track's vocal on the rest; if it doesn't, copy the anchor phrase into every track without changing a character. Which case applies is left as an item for you to check.
Genre comes from two places: what was picked from the list and what you typed yourself. When both exist, what you typed wins, because you typed it wanting something the list didn't have. Where they don't conflict they merge: the picked genre sets rhythm and instruments, the typed words add texture. When both are empty, one is derived from the request with a line on why.
Checklist for writing it yourself
- 1Did you require both the style block and the lyrics block
- 2Do the lyrics contain actual words to sing, not only part tags
- 3Did tempo, key, and instrumentation come as values or recommendations
- 4Did you specify vocal timbre, gender and delivery, and for an album, the track-to-track consistency step
- 5For an instrumental, did you forbid the lyrics field
Reading helps; trying is faster.