Blog

Guide

MiniMax Music 3 Short Song Production Guide: How to Separate Lyrics, Section Tags, and Arrangement Brief

Based on the dual required inputs of the current contract, design a reproducible short song request

Nov 18, 2026
Article cover image: a geometric mineral cross-section rendered in flat facets on a dark teal background, with cyan-green and deep green interlocking, in a restrained style.
On this page

First, Distinguish the Two Required Inputs

MiniMax Music 3 provides only one text-to-music task-based capability on HiAPI; there is no step-by-step interface like "generate melody first, then fill in lyrics." That is, a single request must provide two things at the same time: input.lyrics (required) and input.prompt (required). They are not in a relationship of "one primary, one optional supplement," but are two parallel input channels that jointly determine the output.

Fix the responsibility boundaries of these two fields first, and all later writing issues can be properly placed:

  • lyrics = the content to be sung. It answers "what is sung" and should contain only the lyric body and section structure tags.
  • prompt = the description of the music itself. It answers "what it sounds like" and carries style, mood, vocal texture, instrumentation, and arrangement direction.

Common rework stems from mixing these two up: stuffing "add a drum push in the chorus" into lyrics, or pasting the entire lyrics into prompt. The former causes the lyrics to be sung as sung content or truncated by the section tag rules, while the latter amounts to giving up explicit control over the sung text.

To emphasize the interface form again: this is a task-based interface; after submitting a request, the result is returned asynchronously, and parameters are all under input. Therefore, the recommended writing order is: first determine the lyric skeleton, then write the arrangement brief, and finally use optional parameters to close the length and reproducibility. The next section starts with lyrics.

Two distinct input channels—one for lyrics, one for music description—finally merge into the same generation request.

Correctly Organizing Lyric Text and Section Tags

The rules for input.lyrics are short, but the pitfall rate is the highest: structure tags (such as [intro], [verse], [chorus]) must each occupy their own line; body text written on the same line as a tag will be discarded. This single point explains most feedback about "lyrics lost"—the problem is not the model, but the line breaks.

Basic Form

One element per line: either a tag or a lyric line. Organize a short song skeleton like this:

[intro]
[verse]
Night slowly fades along the windowsill
I fold the words I never finished into the wind
[chorus]
Just let this minute last a little longer
A little longer
[outro]

Three Hard Rules

  1. Tags on their own line. In the form [chorus] Just let this minute last a little longer, the body text after the tag is in a discarded position; the correct approach is to break it into two lines.
  2. Do not add comments to tag lines. Do not put any explanatory text on a tag line besides the tag itself; arrangement intentions should be written into the prompt.
  3. Write lyrics within the same section line by line. Each line corresponds to one natural singing unit; the number of lines itself is part of the rhythm information. Keeping short song sections to a reasonable number of lines makes it easier to get clean results than stacking long sentences.

Organization Template

For a short song, this hierarchy is recommended: [intro] → [verse] → [chorus] → (optional repetition of [verse] / [chorus]) → [outro]. Section tags are structural signals; repeating the same type of tag tells the model "do the same role section again." As for which instruments and what intensity this section actually uses, that is not written here but in the prompt of the next section.

Writing a Layered Arrangement Brief with prompt

input.prompt is a free-text music description. The controllable dimensions specified in the official description include genre, BPM, key, emotional progression, vocal details, and section-by-section arrangement. Rather than writing a large paragraph of adjectives at once, it is better to write in layers from coarse to fine, so that when any dimension needs to be changed, it can be replaced at a specific point instead of rewriting the whole paragraph.

Layering Order

  1. Genre and overall character: one sentence sets the tone, e.g., indie folk, City Pop, piano narrative pop.
  2. Tempo and key: BPM and key are quantifiable anchors; writing them down can significantly reduce randomness.
  3. Emotional progression: describe what emotion the song moves from and to, e.g., "restrained opening, lifted chorus, falling back at the end."
  4. Vocal details: timbre, gender, breathiness ratio, whether to add harmonies—these are vocal-level descriptions.
  5. Section-by-section arrangement: align the section descriptions in prompt with the section tags in lyrics, and specify instrumentation and intensity for intro / verse / chorus respectively.

A Usable Brief Template

Genre: piano-led narrative pop
Tempo and key: mid-slow, minor key
Emotional progression: restrained verse, lifted chorus, closing resolution
Vocals: female voice, close-mic intimate texture, low harmonies added in the chorus
Arrangement: intro with piano and faint ambience only; verse adds bass and light percussion; chorus enters with full drum kit; outro returns to piano

Two Writing Habits

  • Section descriptions share names with lyric tags. If the prompt says "chorus enters full drum kit," then lyrics should have [chorus]; the same names make alignment more direct for the model.
  • Each layer writes only that layer's decisions. Do not mix vocal requirements into the genre line, and do not stuff three mutually exclusive tempos into one sentence. The purpose of layering is to let each iteration touch only one layer.
From the bottom layers of genre and tempo, to emotional progression and vocal details, to the top layer of section-by-section arrangement—the arrangement brief is stacked layer by layer from coarse to fine.

Duration and Random Seed: The Actual Meaning of duration and seed

Both are optional parameters, but they are often misread as "precise control."

duration: An Upper Limit, Not a Contract

input.duration is the upper limit of the generated audio duration, in seconds, default 60, range 1–300. The key point is "upper limit": the model may end early, and the output length is not guaranteed to exactly equal the value you filled in.

This leads to two practical conclusions:

  • For short songs, leave some margin on top of the target length; do not set duration to fit the target duration exactly, otherwise the probability of early ending directly becomes "a chunk short."
  • Do not use duration to control structure. To make the chorus appear earlier, change the section layout in lyrics and the section-by-section arrangement in prompt, not shorten the total duration.

seed: The Anchor for Reproducibility

input.seed takes an integer and is used for result reproducibility; if omitted, a random seed is used. Its value lies in A/B comparison: pair the same lyrics with two versions of prompt, request them respectively under a fixed seed, and differences are easier to attribute to the brief itself rather than to randomness.

Note that seed is only responsible for the semantics of "reproducibility"; it is not a quality knob and does not promise that results under a fixed seed are completely static—but as an experimental baseline for comparing different brief versions, it is the most convenient tool.

Combining into One Short Song Request: Parameter Checklist and Checks

Combining the previous three parts, the parameter field paths for a short song request are fixed:

  • input.lyrics: required, lyric body + section tags each occupying their own line.
  • input.prompt: required, layered arrangement brief.
  • input.duration: optional, duration upper limit, in seconds, default 60, range 1–300.
  • input.seed: optional, integer, used for reproducibility.

Pre-Submission Self-Check List

  1. Does every section tag occupy its own line, with no body text attached to the end of the tag line?
  2. Has any arrangement, instrument, or mood description been mixed into lyrics?
  3. Does prompt cover the five layers: genre, tempo and key, emotional progression, vocals, and section-by-section arrangement?
  4. Do the section names in prompt match the tag names in lyrics?
  5. Does duration leave margin for the target length rather than being treated as an exact value?
  6. If version comparison is needed, has seed been fixed to the same integer?

Going through all six basically avoids most invalid requests. For more detailed fields and calling methods, refer to the minimax-music-3 model page; there are already step-by-step example articles with curl and Python on the site, so they are not repeated here.

Common Failure Modes and Iteration Paths

Finally, list high-frequency problems by "symptom → cause → fix" for easy troubleshooting.

Symptom 1: A Section of Lyrics Is Missing

The most common cause is a tag on the same line as body text, and the same-line body text is discarded. There is only one fix: break [chorus] lyrics into two lines.

Symptom 2: Content That Should Not Be Sung Is Sung

Arrangement requirements were written into lyrics. For example, putting "add drums in the chorus" in the lyric area may cause it to be processed as sung text. The fix is to move all non-lyric content to the section-by-section arrangement layer of prompt.

Symptom 3: The Result Is Shorter Than Expected

duration was used as an exact duration. It is an upper limit, and the model may end early. The fix is to leave margin for the target length and use section layout to drive structure rather than squeezing duration.

Symptom 4: Style Drifts and It Is Unclear Which Version Is Better

No fixed seed and no layered prompt writing mean too many variables between two requests. The fix is to first fix seed, then converge layer by layer in the order "genre → tempo and key → mood → vocals → section-by-section arrangement": in the first round, set only genre and tempo; after confirming the direction, add emotional progression, then add vocal details, and finally fine-tune the section-by-section arrangement. Change only one layer per round, and the difference in any version can be attributed.

Following this path turns short song production from "repeatedly trying your luck" into a convergence process with clear parameters: lyrics manages what is sung, prompt manages how it sounds, and duration and seed manage boundaries and reproducibility. Start with a clean lyric skeleton plus a rough brief, then add details layer by layer.

Related articles

View all articles →

Start building with persistent AI artifacts

Get your API key and call image, video and audio models through one task-based API.

Read the API docs