A Practical Mureka Prompt Guide

Most of what gets written about prompting Mureka is guesswork copied from Suno guides. Mureka publishes its own documentation, and it is unusually specific about what the model takes, so this post sticks to that plus the formatting habits that carry over from any lyric-driven generator.

I want to be upfront about where the line sits. Everything below about Mureka's inputs and limits comes from Mureka's published docs. The lyric-formatting habits come from my own Suno runs, because that is where I have actually burned the credits. I am not going to tell you how a given lyric sheet will sound on Mureka, because I do not know and nobody honestly does.

The two boxes, and what each one is for

Mureka's own quickstart example sends exactly three things to generate a song: the lyrics, the model, and a prompt. That is the whole shape of the job.

The lyrics field is required, and Mureka documents a maximum of 5000 characters. The prompt field is optional and is documented as "Control music generation by inputting a prompt, maximum 1024 characters" in the song generation API. Mureka's own example prompt is four comma-separated fragments: r&b, slow, passionate, male vocal. That is a genre, a tempo feel, a mood, and a vocal direction, and it is worth noticing how short it is.

The model field accepts auto, which Mureka describes as selecting the latest version of the regular model, or a specific version from its published list. Unless you have a reason to pin a version, auto is the sane default.

There are also fields for a preferred vocal gender, a reference track you have uploaded, a cloned voice, and a melody idea. The docs note that a melody reference does not combine with the other control options, while a prompt and a cloned voice do combine.

Easy mode and Custom mode are different jobs

Mureka's Knowledge base describes two ways in. Easy mode is conversational. You describe the idea in a sentence and Mureka writes the lyrics, the title, and the style description itself. The docs say each generation produces two variants and songs are typically two to four minutes.

Custom mode is the one this post is about. It gives you a separate lyrics area with Generate Lyrics and Optimize buttons, and the docs state it supports structured output with verse, chorus, and bridge tags. The documented flow is: write or generate your lyrics, then set Style and Vocal Gender, then hit Create.

If you already have words you care about, Easy mode is the wrong door. It is built to write for you. Custom mode is where you hand over a finished lyric sheet and the style description separately, which is the only setup where the lyric formatting below does anything at all.

The formatting that carries over

Mureka's docs show bracketed section tags in every lyric example they publish. Their quickstart lyrics open with [Verse]. The lyric continuation example opens with [Pre-chorus]. The song extension example opens with [Intro]. Those tags are not decoration in the docs, they are the shape the model's own examples are written in.

The other habit is meter. This one is mine, from Suno, and I am labelling it as such: when two sections that share a melodic slot have wildly different line lengths, the delivery falls apart. The model has to stretch or cram to make the second one fit. There is no reason to expect a different model to be immune to that, because it is a property of fitting words to a repeating melody, not a quirk of one platform.

Before and after

Here is a rough draft the way I actually write them first, all in one block, no tags, lines wandering all over the place.

Before, lyrics:

we drove out past the water tower with the radio down
and I was thinking about everything I never got around to telling you when we still had the kind of time where saying it would not have cost anything at all
the road went quiet
you looked out the window and said nothing for eleven miles and I let you

Before, style prompt:

sad indie song, emotional, good vocals, cinematic, atmospheric, beautiful production

Two problems. The lyric block has no structure at all, so nothing tells the model where a section starts or which lines are the hook. And line two is over forty syllables sitting next to a line of five. The style prompt is worse in a quieter way. Every word in it is an adjective about how good the song should be, and none of them describe an actual sound.

After, lyrics:

[Verse 1]
We drove out past the water tower
Radio down, windows wide
I never said the things I meant to
Back when saying them was free

[Chorus]
Eleven miles of nothing
You looking out, me holding on
Eleven miles of nothing
And the quiet doing all the talking

[Verse 2]
You never asked me where we're going
I never asked you to come back
I kept the exit lights behind us
Back when leaving them was free

After, style prompt:

indie folk rock, mid tempo around 92 bpm, brushed drums, clean electric guitar, upright bass,
restrained male vocal, close and dry vocal mix, no reverb wash

Same song. The verses now sit in the same slot as each other, seven to nine syllables a line, and verse two deliberately echoes verse one's last line so the shape rhymes even when the words do not. The chorus repeats its own first line, which is the lowest-effort way to make a hook read like a hook.

The style prompt changed job entirely. It stopped grading the song and started describing it: genre, tempo, three named instruments, and two things about the vocal. Mureka's own example prompt does the same thing in four words. Fewer adjectives, more nouns.

The order I would work in

1. Write or paste the words first, in Custom mode, not Easy mode.
2. Tag the sections with brackets, using the plain labels Mureka's own examples use.
3. Read the lines out loud and even up the ones that share a slot. This is the step everyone skips and it is the one that changes the most.
4. Write the style prompt last, describing sound rather than quality, and keep it well inside the documented limit.
5. Set vocal gender explicitly if you care about it, since Mureka exposes it as its own field rather than making you bury it in prose.

One more thing worth knowing before you start rerolling whole songs: Mureka documents a fine-tune editing mode that regenerates individual lines the model mis-sang, with an option to keep the existing melody. If one line lands wrong, that is a cheaper fix than a full regeneration.

Why I built Mugical

I ended up writing a tool for myself because doing the tagging and the style prompt by hand, every time, before I could even hear the song once, got old. You give Mugical a song idea, plus genre, mood, instrumentation and era, and it returns a structured lyric sheet with the section tags and ad-lib cues already in place, next to a style prompt covering genre, instruments, mood, tempo and vocal style, ready to paste straight into the style box. It formats for Mureka and for Suno, which want the same two things in the same two boxes.

Evening up the meter between sections is still a pass you do yourself, and so is checking your line endings. That is the part I am working on next.

It does not generate audio. That still happens on whichever platform you paste the output into. What it does is write the structure tags and the style prompt from the same inputs, so they are built to agree, and hand both back ready to paste. The meter stays with you.

You can try it on a song idea and see what the tagged version looks like. If you want to go deeper on just the style side of it, I wrote a separate post on how to structure a Mureka style prompt. And if you also work in Suno and your words keep arriving indistinct, that failure has its own post: why Suno mumbles your words.