Write songs Suno can actually sing.

suno-songwriter is a portable agent skill that turns a request into a release-quality Suno v5.5 package: hook-first lyrics, a clean style prompt, exclusions, slider settings, and section tags that survive generation.

Suno v5.5harness-agnosticno dependencies

What it produces

• Vocal songs: Title, Style, Excluded Styles, and section-tagged lyrics ready to paste into Custom Mode.
• Instrumentals: an Instrumental Prompt plus a cue-style Arrangement map, in the fields Suno's instrumental flow actually exposes.
• Multilingual lyrics: original writing or singable adaptation in the target language, with script and pronunciation decisions stated.
• Style prompts: 200-character style lines built from genre, mood, BPM, vocal identity, instruments, and production texture.
• Personas and albums: reusable artist and album spec files that keep a voice consistent across many tracks.
• Sounds and cover art: one-shot sample prompts and cover art or animated cover prompts for finished tracks.

Design principles

The skill's first priority is a good song, not a feature-complete form. Every vocal lyric passes a quality gate before it ships: the hook lands in the chorus, the chorus is mechanically simpler than the verses, lines stay singable, and at least one concrete human detail anchors the song. A style-diversity rule bans the skill from converging on one recurring palette, so ten requests produce ten different sounds.

It runs on progressive disclosure. A compact SKILL.md handles routing and the core workflow; twenty focused reference files load only when a request needs them. Any harness that can read a skill file and load references selectively can use it.

Quick start

Install the skill folder where your harness looks for skills, then ask for a song.

Install

Download suno-songwriter.zip from the latest release, then pick your app:

Claude Desktop

Settings, then Capabilities, then Skills (Pro, Max, Team, and Enterprise plans with code execution on). Choose Upload skill, select the zip, and toggle it on. Claude triggers it on songwriting requests from the description in SKILL.md.

Codex (desktop app and CLI)

unzip suno-songwriter.zip -d ~/.codex/skills/

Both the app and the CLI read the same directory. Invoke explicitly with $suno-songwriter or let Codex match the request.

Claude Code and other harnesses

unzip suno-songwriter.zip -d ~/.claude/skills/

The skill is plain Markdown with no runtime dependencies; any harness that can read SKILL.md and load references/ selectively can use it.

First requests

Use $suno-songwriter to write a modern country song about a father
driving through the night to pick up his kid.

Use $suno-songwriter for a no-vocal lofi study beat with rain,
Rhodes, and soft drums.

Use $suno-songwriter to write a Japanese hyperpop chorus about
surviving a terrible night. Native script.

Use $suno-songwriter to improve this Suno v5.5 prompt.

What comes back

A complete package in the output contract: paste the Style line into the Styles field, the Excluded Styles line into Excluded Styles, the lyrics body into Custom Mode lyrics, and set the sliders as suggested. Optional sections (Language Notes, Suno v5.5 Notes, Checks) appear only when they earn their place.

Output contract

Two shapes, matched to the two Suno flows. Fields never leak between them.

Vocal songs (Custom Mode)

Instrumentals (Instrumental Mode)

Suno's instrumental flow does not expose Style, Excluded Styles, or sliders, so instrumental packages never include them:

Workflow & quality gate

The skill writes a song first and a prompt second. The form exists to serve the record.

Core workflow

1. Clarify the goal: theme, genre, language, vocal or instrumental, audience, emotional arc. Saved personas and albums load first and act as constraints.
2. Build the song as a song: title, central promise, hook phrase, section map, then lyrics.
3. Treat supplied lyrics as reference material unless a close edit is requested.
4. Choose the sound from the request; never default to a recurring palette.
5. Keep Suno fields clean: positive direction in Style, negative in Excluded Styles, cues in the lyrics body.
6. Run one silent revision pass before output.
7. Add optional sections only when they earn their place.

The lyric quality gate

Every vocal lyric must pass before it ships:

• The title or hook phrase appears in or near the chorus.
• The chorus is shorter, plainer, and more repeatable than the verses.
• Lines are compact and singable, usually 5 to 9 words, one image or claim each.
• Point of view and tense stay stable unless a shift is deliberate.
• At least one concrete human detail anchors the song.
• The final chorus, bridge, or outro shifts the meaning slightly instead of repeating the same emotion.

Anti-generic rules

Stock filler words (neon, flames, shadows, echoes, fading, forever) are rewritten unless the user wants them. Strong genre archetypes stay when they serve the hook. Living artists are never named in the Style field: the skill translates a reference into era, scene, and production traits instead.

Lyrics & structure

Suno-native phrasing by default, with section tags that steer the performance.

Default shape

• 4-line verses, 2-line pre-choruses, a 4-to-6-line chorus with one unmistakable hook.
• Plain syntax: subject, verb, image. One strong image can carry a section.
• Familiar singable rhyme families over forced cleverness.
• Sections stay at 4 lines or a multiple of 4 unless the genre says otherwise.

Groove-narrative mode

For neo-soul, R&B storytelling, jazz-hop, conscious rap, and city-vignette songs, verses can run long (8, 12, 16 lines) over a steady groove, with tight rhythmic couplets, vocalist labels like (Male) and (Female - wordless hum), environmental cues in the arrangement tags, and spoken bridges. The chorus still stays simpler than the verses.

Section tags and cues

[Verse 1]           [Whispered Verse]        [Guitar Solo]
[Pre-Chorus]        [Explosive Chorus]       [Beat Drop]
[Chorus]            [Stripped Bridge]        [Male Voice] / [Duet]
[Bridge]            (ad-libs in parens)      [End] / [Fade Out]

Short performance descriptors inside section tags cue dynamic changes. Persistent production direction belongs in the Style field, not in tags. [End] after the final section is the reliable way to stop Suno inventing extra verses. Too many tags get ignored, or worse, sung aloud.

Style & genres

One 200-character line does most of the sonic steering. The skill builds it deliberately.

The style line formula

[Genre/subgenre], [mood], [tempo/BPM], [vocal style],
[instruments and production], [optional theme]

garage rock, restless and defiant, 132 BPM, raw group vocals,
distorted guitars and live drums, punchy basement production

Anti-lock-in

Before a style line ships, the skill asks whether it was chosen for this request or copied from habit. If copied, at least three dimensions change: genre, BPM, vocal identity, rhythm section, lead instruments, production texture, or structure. A palette table covers 45+ lanes, from boom bap to bossa nova to trot, with BPM bands, signature instruments, and structures used as a menu, never as defaults.

Reference artists

"Make it sound like [artist]" never passes the name through: Suno's filters can reject it or flatten the result. The skill decodes the reference into era, scene, rhythm section, signature sound, and production character. late-90s East Coast boom bap, dusty chopped soul sample, knocking drums travels; a name does not.

Excluded Styles

Short and specific: unwanted genres, instruments, vocal traits, or production moods. Never excludes what Style asks for, and never used to fix lyric problems.

Genre fusion

Three different requests hide behind "mix these genres". The skill routes them explicitly.

1. Fusion: one hybrid sound

Built as anchor + accent, never 50/50. The anchor genre owns the rhythm section, tempo, and structure; the accent contributes two or three signature elements. BPM comes from the anchor's band. Established fusion names (folktronica, country trap, electro swing) beat invented pairings, and Excluded Styles blocks the generic middle the blend could collapse into.

cumbia surf rock, salty and playful, 102 BPM, twangy reverb guitar
over güira and congas, relaxed male group vocals, dry garage mix

2. Mid-song genre switch

The arc is stated in the Style line in natural language and pinned with a section tag: [Chorus - double-time house flip], [Beat Switch]. One or two constants (vocalist, key, a motif) hold across the switch so it reads as one song. One switch per song is the sweet spot. If takes keep averaging the two sounds, split-generate and join with Extend.

3. Genre versions of one song

Same title and lyric core, separate labeled packages, each with its own Style, Excluded Styles, and sliders. Singability is re-checked per tempo: a lyric metered for a 74 BPM ballad will not scan at 140 BPM drill, and the changes are stated, never silent.

Instrumental & Sounds

Cues, loops, beds, and one-shot samples, in the fields Suno's flows actually expose.

Instrumental tracks

Lofi, OST cues, game loops, ambient, trailer beds, and background music return an Instrumental Prompt: plus an Arrangement: cue map. No Style, Excluded Styles, or sliders. Loopable tracks avoid hard endings and mark [Loop Point]; cues describe motif, dynamics, and the emotional turn.

instrumental lofi chillhop study beat, relaxed 78 BPM, dusty drums,
mellow Rhodes, soft upright bass, vinyl texture, loopable, no lead vocals

"Instrumental" does not always mean voiceless: wordless chops, choir aahs, and chant hooks are described in the prompt when wanted, and fully excluded (no vocal chops, no chants) when not.

Suno Sounds (samples)

Individual samples (a whoosh, a riser, an 808 kick, an ambience bed) are Suno's Sounds feature, not a song. They return a different block: Sound: prompt, Type: (One Shot or Loop), BPM: and Key: for musical material only, and Variations:. Duration is stated in the prompt when it matters: 5-second sci-fi teleport swoosh.

Multilingual

Original writing, singable adaptation, or deliberate code-switching, with the language decisions stated instead of assumed.

What gets clarified

• Target language and dialect: Spanish (Mexico), Arabic (Levantine), Japanese, Korean.
• Native script, romanization, or both.
• Original lyrics vs singable adaptation vs literal translation vs bilingual rewrite.

Structure follows the language

English pop density does not transfer. Romance languages need more syllables per idea; Japanese and Korean pop tolerate compact phrases and English hook accents; Spanish hooks favor open vowels over exact rhyme. "Korean music" spans ballad, K-hip hop, K-R&B, indie, OST, and trot, not just K-pop, and the skill asks or infers which lane.

Pronunciation

AI vocal models anglicize vowels and mis-stress unfamiliar words. The skill plans for it: in-line phonetic respelling for words that reliably break, a separate Pronunciation Notes: block otherwise, and a native-speaker review recommendation whenever release quality matters.

Screen music

Anime OP/ED, TV themes, trailer cuts, idents, and game themes: music tied to picture and to an edit length.

Routing by format

Hitting the cut

Edit lengths are hit with the Duration control and a structure written to size: set Duration 5 to 15 seconds longer than the target so the ending resolves (about 1:40 for a TV-size ~90s OP), cap with [End], then Crop to the exact time. The slider alone will not compress an oversized lyric. The hook or motif must land inside the first 30 seconds, sooner for short formats.

Style match & remix

Style yes, expression no. The skill imitates an approach, never a recording.

The hard rule

Cadence, rhyme strategy, structure, subject matter, and genre are free to write like. Specific lyrics, hooks, and distinctive phrases are not: never reproduced, closely paraphrased, or continued. Artist names stay out of the Style field, both for the legal line and because Suno's filter can flatten a take that trips it.

Style-match workflow

1. Fingerprint the source: cadence and flow, rhyme density, line shape, diction register, POV and themes, and the sonic lane.
2. Sharpen with search when available, pulling descriptions of the style, never verbatim lyrics.
3. Generate original work on a fresh subject so the result is its own song, not a re-skin.
4. Declare it: a Style Reference: line names what was modeled and states that all lyrics are original.

Remix mode

For lyrics the user owns, close rework is fine: the skill states in one line what is kept (premise, arc, a motif) and what changes (genre, structure, images), then rewrites. Ownership gets confirmed when it is ambiguous, and requests for actual copyrighted lyrics are declined with a style-match offered instead.

Personas & albums

Consistency across many tracks, stored as small spec files in your project, not in the skill.

Personas

A persona is a reusable artist: identity, vocals, cadence, lyrical fingerprint, sonic palette, excluded styles, and Suno defaults, saved at personas/<slug>.md in your working directory. Later requests like "write this as [persona]" load the file and treat it as a constraint. Style-reference personas decode a real act into a fingerprint and keep the name out of every prompt; the persona is a fictional act with your chosen name.

Albums

An album spec at albums/<slug>.md adds a concept, a shared Style kernel, cohesion rules (what stays constant vs what varies track to track), recurring elements, and a sequenced tracklist. Writing "track 4 of [album]" builds the song as the album's kernel plus that track's mood and theme, then updates the tracklist and the persona's song log in the same pass.

Controls & sliders

What each Suno v5.5 input does, and how the skill sets it.

Creative sliders

Always ranges, never false-precise single values. Troubleshooting moves in 5-to-10-point steps, one variable at a time.

Duration

v5.5's Create form has a Duration control: Auto (the model picks a natural length) or Custom, a slider from 0:10 to 6:00 in 5-second steps. Custom is for target cuts: TV-size themes, stings, loops of known length. The slider does not compress the lyric, so the structure is written to fit the time, Duration is set slightly long so the ending resolves, and Crop trims to the exact cut. The reliable band for full songs is roughly 2:00 to 3:30.

Personalization

• Voices: with a trained Voice selected, the skill stops over-describing the vocal and trusts the Voice.
• Custom Models: prompts aim at song intent instead of fighting the model with contradictory genre tags.
• My Taste: the Styles-field Magic Wand seeds personalized ideas; the skill then edits for specificity.

Content filters

Real-person names, brands, heavy profanity, and graphic content can make Suno silently reject a take or return something oddly bland. When a generation comes back flat, the skill suspects the filter before the sliders: soften or imply the offending material, and note the change so you know why a line moved.

Studio & iteration

Getting from a promising take to a finished record.

The toolkit

Edit vs re-roll

The skill's rule of thumb: fix a section, keep a take; fix a song, change the prompt. Slider changes move in small steps with the prompt held constant, so you can tell which change did what.

Reference index

The skill loads these on demand. Each file owns one topic.

Changelog

What changed, most recent first.

v1.2.1 (2026-07-31)

• Changed: prose quality pass across SKILL.md and all 21 references (dashes rewritten, taught tag syntax unified on plain hyphens, filler vocabulary replaced); added a KIKO.md voice spec. No behavioral changes.

v1.2.0 (2026-07-31)

• Added genre fusion: anchor + accent blends, mid-song genre switches, and multi-genre versions of one song, with smoke tests.
• Added a groove-narrative vocal mode for neo-soul, R&B, and spoken-sung storytelling.
• Added release packaging (the install zip), per-app install instructions, this docs site, and the MIT license.
• Fixed the lyric quality gate: hook-first priorities, simpler choruses, singable line lengths.
• Fixed Instrumental Mode guidance to use only the fields Suno's instrumental flow exposes.
• Documented v5.5's Duration control (Auto, or Custom 0:10 to 6:00 in 5-second steps).

v1.1.0 (2026-05-12)

• Added the cover art and animated covers reference.

v1.0.1

• Shortened the SKILL.md description and dropped XML-style tokens for harness compatibility.

v1.0.0

• First tagged release: core workflow, output contract, and the initial reference set.