When a generated track misses the brief, the first diagnosis is usually the model. A more testable diagnosis is the input. If the request was one paragraph mixing a story, a genre, a celebrity reference, and a private joke, you cannot tell whether the system failed a requirement or never received one.

This is an interface-level playbook, not a source-code analysis and not a listening test I ran. The aim is to write the request so a later take can be judged against fields, not against a vibe.

A paragraph is not a job definition

Text-to-song systems look conversational, so people write to them conversationally. That is reasonable for exploration, and a poor format for a job you need to revise.

A job definition for this class of tool usually has at least five parts, whether or not the UI shows them as separate fields:

  • the use of the track (bumper, full listen, under speech, personal draft)
  • the lyric path (locked words, system-written words, or instrumental)
  • a style contract (sound, not a biography)
  • any hard limits (duration, vocal language, exclusions)
  • the conditions under which the file may be used

If those parts are mashed into one sentence, a second run cannot change one variable, and you cannot write a useful failure note. "It doesn't sound right" is not reproducible. "The locked chorus was not placed as a chorus" is.

The style contract: four fields, then limits

A practical style contract can start with four named fields:

  1. Genre, or a tight genre blend
  2. Mood
  3. Two or three instruments
  4. Vocal character, or an explicit instrumental flag

A style line such as "dreamy synthwave lullaby, soft analog pads, breathy female vocal, 90 BPM" can be edited field by field. "Make it sound like my favorite radio single, but sadder, and a bit like a film trailer" cannot.

Add limits next to the four fields, not inside the lyric:

  • target length, if the tool accepts a duration
  • language of the vocal, if that matters
  • styles or textures to exclude, if the tool has an exclusion control
  • male, female, auto, or instrumental, if those are explicit states

Do not put a real artist name in the style contract. Some systems reject those prompts. Even when they do not, "in the style of X" is weaker and legally messier than era, tempo, instrument, and vocal tone.

Do not store the lyric in the style contract. The style contract describes the record. The lyric map, in the next section, describes the words and the sections. If a tool gives you two boxes, use them as two boxes.

The lyric map: tags, a short hook, and singable lines

When the words are yours, the input is a map, not a poem dump.

Mark sections on their own lines with tags the system documents, such as [Verse], [Pre-Chorus], [Chorus], and [Bridge]. The tags are how you state where the hook lives. If you do not mark a chorus, you are asking the model to invent the architecture and then blaming it for a flat arrangement.

Keep the chorus short, and give it one line that can repeat. Long choruses spread the hook until it stops behaving like a hook. Verses can carry detail; the chorus should stay sayable.

Read the lines out loud before you generate. A clause that looks precise on a screen may be too dense to sing. Unusual names, model numbers, and digits are frequent diction failures. When a word is sung oddly, a common published fix is to spell it the way it sounds and write numbers as words, then generate again before rewriting the lyric or switching models.

If you do not have lyrics yet, that is a different schema: send a story or feeling to a lyrics draft, then promote the approved text into the lyric map. Do not leave the lyrics box empty and later score the take as if the words were locked; depending on the model, an empty box may be a request to write words.

Record published constraints before you spend a run

Once the style contract and lyric map are stable, choose an interpreter with its published constraints in view. This is a pretest log, not a ranking.

Useful questions, all answerable from a public interface or help text:

  • How many takes does one run return?
  • Can I set length, or is length automatic?
  • Will this model sing pasted lyrics, write lyrics if the box is empty, or favor instrumentals with a duration control?
  • What does a run cost in the system's credits, and is that cost visible before I generate?
  • What file formats can I export, and under which plan is commercial use actually granted?

SongMaker is a public example of a workspace that hosts five models — Suno, Mureka, ElevenLabs, Lyria, and MiniMax — and invites you to run the same prompt and keep the take you prefer. That matters for the schema approach: the contract can stay fixed while the interpreter changes. The site describes a path from a text prompt or pasted lyrics to a full song, with vocals or instrumental. Extend and remix are outside this playbook. I am not claiming a result from using them.

Those homepage facts still leave input fields underspecified. A schema is only useful if you can point to the controls that correspond to each field, and to the published limits that belong in the pretest log.

The AI Song Generator page publishes a more detailed contract surface: Simple versus Custom mode; style tags plus free text; an exclude-styles control; male, female, auto, or instrumental; vocals in 13 languages; two takes per run on some models; section tags, including a "/" shortcut to insert a tag; synced lyrics playback; and download formats listed as MP3, WAV, or FLAC. Published notes you can copy into a pretest log include: Suno at 5 credits per run, two songs, 10 seconds to 6 minutes in Custom mode; Mureka from 5 credits per run, two takes; ElevenLabs at 12 credits per minute, up to 5 minutes; Lyria at 6 credits per run, one song, described as singing pasted lyrics in Custom mode; MiniMax at 12 credits per run, and able to write lyrics when that is the job. New accounts start with 10 free credits. On licensing, the page is explicit in a way the schema should copy: a commercial license is included with annual plans; Free and monthly plans are for personal use. YouTube, Spotify, and similar platforms have their own rules for AI-generated music.

Treat those figures as published claims to re-check on the live page before you encode them into a runbook. They are not benchmarks, and they are not my measurements.

A log that stays honest

A useful run log records facts, not invented quality scores:

  • the style contract, field by field
  • the lyric map, including tags
  • the model and the published constraints you accepted
  • the questions you will ask on first listen (hook placement, lyric fidelity, diction, vocal character, length)
  • what you will change on the next run — one field, not five

If you change the lyric, the style, the length, and the model at once, you no longer have a test. You have another paragraph.

The limitation of this playbook is the same as its point. A clean schema will not make a weak idea strong, and it will not tell you how a distributor will treat a track. It will only make failure legible: the difference between "the model is bad" and "the chorus was never marked," which is a difference you can act on.