The five-part style formula
Nearly every bad Suno result traces back to the same thing: a Style box that reads like a mood, not a brief. "sad piano song" gives the model four hundred equally valid directions and it picks the most average one. The fix is to fill five slots, in this order, because Suno weights the earliest terms most heavily.
rock is a cloud; garage rock revival is a sound. Put the dominant genre first when you're blending - it wins the ties.desperate outperforms minor key with reverb, because the model has heard far more music described that way.raspy female), delivery (breathy, close-mic'd) and treatment (doubled, tape-slapped).clean jangly Telecaster and brushed drums carry tone; guitar and drums carry none. Two or three lead voices is plenty.lo-fi tape warmth, 92 BPM anchors the groove; without a number the model drifts toward whatever the genre averages.Style prompt builder
Pick a value for each slot and watch the prompt assemble in the right order. Copy it, paste it into Suno's Style box, then edit by hand - the point is the shape, not the vocabulary.
Where each instruction belongs
The single most common Suno failure is an instruction in the wrong field - and the punishment is specific: anything you type in the Lyrics box that isn't a bracketed tag gets sung out loud. People have shipped songs where a voice sincerely sings "upbeat pop, female vocals, 120 BPM".
| Put this | In this field | Because |
|---|---|---|
| Genre, instruments, production, tempo | Style | It describes the whole song's sound world. |
| Overall vocal character and delivery | Style | It applies from the first bar to the last. |
| The words you want sung | Lyrics | Everything in this box is treated as performable text. |
[Verse], [Chorus], [Bridge] | Lyrics | Section tags organise the arrangement, and aren't sung. |
| A cue for one section only | Lyrics | e.g. [Final Chorus - gospel choir, full intensity]. |
| Things you don't want | Exclude Styles | Negations inside the Style box are unreliable. This field is the real control. |
| Song length | Duration slider | Asking in prose does nothing. The slider is the knob. |
| How literally to follow you | Style Influence | Adjectives can't set adherence. The slider can. |
Suno meta tags, by job
Meta tags go in the Lyrics box, in square brackets, on their own line. They're guidance rather than commands - the model honours them most of the time and ignores them when they fight the style - but structure tags are the difference between a song with a shape and four minutes of verse. Click any tag to copy it.
Structure the ones that always earn their place
Vocals who is singing, and how
Instruments hand the spotlight to something
Dynamics shape the energy between sections
Stacked tags beat comma-separated ones
A bare [Chorus] is the most reliable tag there is - and also the most likely to override modifiers you tack onto it with commas. Stacking with pipes reads more cleanly to the model, and section-scoped cues after a dash are honoured more often than the same words buried in prose.
[End]. It's the single most useful tag on the list - roughly nine times in ten it stops the trailing noodle where the model keeps playing after the song is over.
[Chorus] tag. Identical lyrics are the strongest signal you can send that the same melody should come back.
Writing lyrics Suno can actually sing
The Lyrics box holds 5,000 characters on v4.5 and up, but filling it is a trap. Somewhere around 3,000 characters - roughly 40 to 60 lines - the model starts rushing to fit everything into the runtime, and phrasing goes flat. Say less and it sings better.
Section lengths
Verses: 4-8 lines. Choruses: 2-4. Longer verses get compressed into a mumble; longer choruses lose the hook. Keep line lengths within a syllable or two of each other so the meter holds.
Adherence fades
Prompt influence decays after the first 30 to 60 seconds of a generation. Front-load what matters, and put the section that has to be right early - or build the song in pieces with Extend.
Mood has to match
A grief-stricken lyric under a euphoric festival house style gives you neither. The Style box and the words need to want the same thing, or the model splits the difference and lands nowhere.
The sliders that do what adjectives can't
Under Advanced Options sit three sliders and one field that most people never open. They're blunter and more reliable than any wording you can put in the Style box - and using them frees up characters for the things prompts are actually good at.
At the bottom you get textbook, commercial, slightly anonymous. In the middle you get songs that sound like a person made choices. Past about 80% the model starts fragmenting on purpose - useful for sample fodder, useless for a full track. 40-60% is the right default for an original song.
Low treats your tags as hints. High treats them as instructions. Pair a high setting with a lean 5-8 tag prompt - crowded tag lists at maximum adherence produce a mix where everything you asked for is present and nothing is audible.
How much an uploaded clip drives the result. Around 55% is the sweet spot: the melody survives while the arrangement gets fully re-rendered. Treat this and Style Influence as competing forces - push one, ease the other. 30-60 seconds is the useful sample length.
Exclude Styles: the negative prompt that works
Typing "no drums" into the Style box mostly teaches the model the word drums. The Exclude Styles field is the documented control, and it holds a handful of comma-separated terms. Keep it to your two or three most important exclusions - a long list confuses the model as surely as a long prompt does.
| The problem | What to put in Exclude Styles |
|---|---|
| A vocal keeps appearing on an instrumental | vocals, singing, choir, vocal chopsBelt and braces: also add [Instrumental] in Lyrics. |
| It went to a female voice and you wanted male | female vocals |
| Everything drifts toward EDM | EDM, build-up, drop, sidechain |
| Trap hats sneak into your boom-bap | trap, hi-hat rolls, 808 slides |
| Vocals arrive drenched in autotune | autotune, pitch correction |
| The mix is a wash of reverb | reverb, ambient pads, wide stereo |
| An acoustic song keeps growing synths | synthesizer, synth pads, electronic drums |
| It turns into a power ballad at the chorus | orchestral swell, key change, stadium drums |
40 Suno prompts that work
Every one of these follows the five-part order and lands between eight and fifteen tags. They're starting points, not answers - swap the instrument or the BPM and you've got a different song. Click the icon on any prompt to copy it.
Pop & synth-pop
Radio-shaped songs with a hook that has to land by 0:45.
Hip-hop, trap & R&B
Beat-forward material where the pocket matters more than the melody.
Rock, punk & metal
Guitar-led work where the amp and the room are half the sound.
Folk, country & acoustic
Songs that live or die on the vocal being believable.
Electronic & dance
Where the arrangement is the song and the vocal is a texture.
Cinematic & orchestral
Score-shaped pieces for trailers, games and video beds.
Lo-fi, ambient & focus
Long-form background material where nothing should demand attention.
Jazz, soul & funk
Performance-led styles where feel beats precision.
Why your songs sound generic
Almost always one of these six. Note the pattern - every fix is about giving the model something concrete to hold on to instead of something it has to guess.
What people write
- "pop song, upbeat, catchy"Three words the model can satisfy a thousand ways. It picks the average.
- "realistic, professional, high quality"Ignored. These describe your hopes, not a sound.
- "sad but also hopeful and angry"Competing directions cancel out into flat.
- "female vocals"One layer out of three. No character, no delivery, no treatment.
- "guitar, drums, bass"Instrument nouns with no tone attached.
- No tempo at allThe model defaults to whatever the genre averages.
What to write instead
- "synth-pop, indie pop"Two levels deep names an actual sound, not a category.
- "close-mic'd with audible breath, room tone left in"Describe the physical recording and you get realism.
- "defiant"Pick one direction; use
[Bridge]for the contrast. - "raspy female vocal, breathy delivery, doubled and tape-slapped"Character, delivery, treatment.
- "clean jangly Telecaster, brushed snare, melodic Rickenbacker bass"Named gear produces distinct tone.
- "104 BPM"One number that anchors everything else.
Personas, Voices and Custom Models
v5.5 added three ways to carry an identity between songs, and each one changes how much prompt you should write. The rule is the same in all three cases: the more identity the system already holds, the less your Style box should try to describe.
Personas
Made from one of your own songs, a Persona carries its vocal character, genre, instrumentation and production texture forward. It doesn't lock an identical singer or melody.
Prompt accordingly: Suno auto-fills the Style box from the source. Delete anything that contradicts it - a new era, a different singer, a clashing production note - or the Persona drifts.
Voices
Record or upload your own voice and sing on your own tracks. Verification requires reading a displayed phrase aloud. Requires v5.5, and availability is still expanding by region.
Prompt accordingly: drop the vocal-timbre tags entirely - the voice is already decided. Spend those characters on delivery and arrangement instead.
Custom Models & My Taste
Upload at least six tracks from your catalogue and v5.5 trains a version of itself that knows your sound. My Taste quietly learns your go-to genres and applies them through the Magic Wand.
Prompt accordingly: keep prompts short - the model already holds your defaults. Be aware that My Taste pulls everything toward a house sound over time.
Suno prompt FAQ
How long should a Suno prompt be?
The Style box holds 1,000 characters on v4.5 and up, but the useful range is 8 to 15 tags - typically 150 to 300 characters. Under five tags is vague; past twenty they start contradicting each other.
What are the character limits?
On v4.5, v5 and v5.5: 1,000 for Style, 5,000 for Lyrics, 100 for the title. On v4 and earlier it was 200 / 3,000 / 80. The limit follows your model version, not your plan.
Does the order of tags matter?
Yes, considerably. The first tag carries the most weight. When you're blending two genres, whichever you put first wins the ties - lead with the one that should dominate.
Why is Suno singing my prompt?
You put it in the Lyrics box. Everything there that isn't inside square brackets is treated as words to perform. Move genre, instrument and production descriptions to the Style field.
How do I get an instrumental?
Use the Instrumental toggle. If a vocal still creeps in, stack three defences: no vocals in the Style box, [Instrumental] in the Lyrics box, and vocals, singing, choir in Exclude Styles.
Can I name a real artist in a prompt?
Suno filters requests that name living artists, and it isn't a good idea in any case - you'd be building a release on a similarity claim. Describe the attributes instead: era, production style, vocal character, instrumentation. You get closer, and it's yours.
How do I stop the song trailing off?
Put [End] on the last line of your lyrics. It's about 85-90% effective at cutting the outro noodle. The Duration slider handles overall length; asking in prose does nothing.
Do meta tags actually work?
Structure tags do, reliably. Modifiers attached to them are guidance - [Chorus] is the strongest tag on the list and also the most likely to override extras you hang off it. Pipe-stacking ([Chorus | Anthemic | Full Band]) holds better than commas.
Why does the second half sound different?
Prompt adherence decays after the first 30-60 seconds of a generation, and mid-track drift toward generic pop is a known v5.5 behaviour. Front-load the sections that matter, or build the song in stages with Extend.
What should Weirdness be set to?
40-60% for most original songs. Below 20% is textbook and commercial; above 80% the model deliberately fragments, which is only useful if you're farming samples.
Do prompts cost credits?
Writing them doesn't - generating does, every time. Budget roughly 50-100 credits for a track you're happy with once iteration is counted. Spending five more minutes on the prompt is cheaper than five more generations.
Is there an official prompt generator?
The Magic Wand inside Suno expands a short idea into a full style prompt, and on v5.5 it draws on My Taste while doing it. It's a good first draft - then edit it into the five-part order yourself.
Write the spec, not the wish
Copy a prompt from the library, change one thing, and listen to what moves. Ten minutes of that teaches you more about Suno than a hundred regenerations of "upbeat pop song".