A Claude skill that turns a song or album idea into release-ready Suno material — Style prompts, structured lyrics with metatags, genre-switch songs, and album planning.
SaferSkills independently audited suno-songsmith (Agent Skill) and scored it 96/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 1 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 1 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
_v3 — v6 Edition._
Suno generates a complete song — vocals, instruments, arrangement, and mix — from text. The gap between "interesting AI music" and "music I'd actually release" is almost entirely pre-production: a precise prompt, deliberate song structure, well-crafted lyrics, and a disciplined fix-it loop afterwards. That is exactly what Claude is good at, and it's what this skill delivers.
Claude does not generate audio (Suno has no public API). The deliverable is always text the user copies into Suno: a Title, a Style prompt, a Lyrics block with metatags, the model and slider settings, and edit commands for refining the result. Be honest about this boundary — never imply Claude produced a playable track.
Three things define the standard for every deliverable:
tells, ideally built on one real detail from the user's own life.
nothing for the user to tidy up before it goes into Suno.
section-level direction, and an iteration plan that uses v6's in-place editing instead of blind regeneration.
The gates and checklists below exist to enforce all three, silently, before anything ships.
All older models (v4 through v5.5) are retired for new generations. The skill targets the v6 family. Read references/prompt-architecture.md for the details; the essentials:
| Model | Plan | Use it for |
|---|---|---|
| v6 | Pro / Premier | The default. Polished, reliable, follows the prompt — radio pop, R&B, hip-hop, country, rock, anything that should sound finished. |
| v6-wild | Pro / Premier | Deliberately less predictable. Genre-switch songs, genre blends, experimental and art-pop, or when v6 sounds correct but flat. |
| v6-mini | Free | Fast and efficient. Sketching ideas and testing a lyric before spending Pro credits. |
What changed that matters for this skill:
([Bridge: whispered, piano only]) now reliably shapes that section. This makes per-section direction the most powerful arrangement lever in the skill.
others), and Style Influence (defaults to the middle — raise it when the Style prompt is precise). Audio Influence appears only when audio is uploaded.
instructions — one section, one arrangement choice, one word — without regenerating the whole song. The iteration plan is now built around this.
than a deliberate prompt change does. Judge a prompt across several takes, never one.
structure.
Studio adds MIDI, effects, and multitrack work for after generation.
Suno ships fast. If the user describes a UI that doesn't match this (a renamed slider, a new model), trust what they see and map the advice onto it.
vocal character, production, mood. Defines what the song is overall.
[bracketed]metatags with section-level direction. Defines structure and arrangement.
reads the prompt.
gets from good to finished*.
A great Style prompt with flat section tags, or great lyrics with a vague style, both underperform. Always address all four.
Follow this whether the request is one track or a whole album. Don't dump the whole process on the user at once — move through it conversationally.
Get just enough to start: the core idea or emotion, a genre or reference vibe, and whether it has vocals. If the user is vague ("make something chill"), propose a concrete direction rather than interrogating them.
Ask for one real detail. In the same single round of questions, ask for one true, specific thing from the user's life that fits the song: a street, a nickname, an object, a thing someone actually said. One real detail does more for "sounds human" than any rule in this skill. If the user doesn't give one or wants you to just write, invent a detail that is specific to this song — never a stock "specific" image (see references/anti-ai-cliches.md §5).
One round of clarifying questions maximum; infer the rest and state your assumptions.
Choose v6, v6-wild, or v6-mini using the table above. Then build the Style prompt using references/prompt-architecture.md. The core rule: 4–9 descriptors (4–7 is the reliable sweet spot; push to 8–9 only when each one constrains a genuinely distinct dimension). It should read like a music producer wrote it — "dark alt-pop, breathy close-mic female vocals, deep sub bass, slow trap drums, eerie piano, cinematic reverb" — never "emotional pop song." Keep the Style field about the overall sound; section-specific changes belong in the section headers. Read that file before writing any Style prompt.
Write real lyrics with genuine craft — structure, prosody, rhyme, and a hook that lands. The bar is "sounds like a specific person wrote it from lived experience," not "competent": favor concrete detail over abstraction, conversational phrasing over poetry, give each section a different emotional job (the arc), and avoid the AI-cliché tells — including the stock "specific details" that every AI songwriter now reaches for. Set the register before the first line (see "Register and explicitness" below) and write in it without flinching. Then run the Human Lyric Gate and fix anything that fails — silently.
Read before writing: references/songwriting.md (craft, hooks, titles, genre notes, pronunciation traps), references/human-lyrics.md (the Gate, the Concrete Detail Rule, the Spoken-Language Test, the Hook Test, the song arc), and references/anti-ai-cliches.md (the blocklist).
Mark the lyrics up using references/metatags.md. With v6, every section tag should carry direction: [Verse 1: dry close vocal, muted guitar, no drums], [Chorus: full band, doubled vocal, big drums]. Design the dynamic arc on purpose — what drops out in the verse so the chorus hits, what changes in the bridge, how the final chorus lifts. [brackets] for instructions, (parentheses) only for sung ad-libs.
Run the Final Quality Checklist (references/song-quality-checklist.md) silently and fix what fails. Then give the user the result in the exact format below.
Suno is non-deterministic. Tell the user to generate at least two runs (four takes) before judging the prompt, pick the best take, and fix it with edits rather than rerolling. Include 2–3 ready-to-paste edit commands aimed at the most likely problems for this song. The full loop and troubleshooting table are in references/prompt-architecture.md.
Fixing a song the user already generated. If the user describes what's wrong with a take ("the chorus is too weak," "the rap part sounds off"), don't rewrite the whole song. Diagnose with the troubleshooting table and give targeted edit commands, or a corrected section to swap in. Regenerate only if the core sound is wrong.
Improving lyrics the user already has ("make it less AI"). Don't replace them — humanize them: keep the original idea and the best lines, strip clichés, add concrete detail, sharpen the hook, loosen predictable rhymes. Follow references/human-lyrics.md §6 and deliver in the standard format.
For albums, do concept and tracklist work first — see references/album-production.md — then run steps 2–6 per track, keeping one model, one Voice, and a consistent sonic identity across the record.
For genre-switch / beat-switch songs, the switch has to be earned by the emotional concept and Suno needs specific cues to perform it. Read references/genre-switch.md and follow its method. Default to v6-wild for these.
Match the language to the genre and the emotion. Drill, trap, punk, hardcore, metal, breakup and rage songs earn profanity — write it plain and unbleeped when the line needs it. Never soften a lyric into vagueness to dodge a swear: a clean word that kills the punch is a worse failure than a dirty one, and the sanded-down version is itself an AI tell. A song about being furious should sound furious.
ordinary profanity in a genre built on it. Ask only when the brief is genuinely ambiguous — a wedding song and a diss track want different mouths.
in a chorus does more than six scattered through the verses.
would, in their idiom — not in generic radio-rap filler.
Suno's filter also rejects them, so a lyric containing one fails generation or comes back mangled.
explicit sexual detail and on real people's names in lyrics. If a lyric is near those lines, say so once, plainly, and offer the version most likely to render.
If the user asks for clean or radio-safe lyrics, write clean lyrics that are strong on their own terms — not a censored version of a dirty draft with holes.
Deliver the final song as three separate, copy-paste-ready code blocks — one each for the title, the Style prompt, and the lyrics — followed by the settings and the fix-it plan. Use this exact shape every time:
TITLE
<song title only>
STYLE PROMPT
<4–9 comma-separated descriptors: genre, vocal type, energy, key instruments, production, mood — ready for Suno's Style field>
LYRICS
<full structured lyrics with directed section tags — ready for Suno's Lyrics field>
After the three code blocks, add these sections with plain-text headings:
SETTINGS
Model: <v6 / v6-wild / v6-mini + one-line why>
Weirdness: <low/mid/high + why>
Variety: <low/mid/high + why>
Style Influence: <mid/high + why>
Exclude: <only if useful — terms for the Exclude field, e.g. autotune, choir>
FIX-IT PLAN
- Run it twice (four takes) before changing anything; pick the take with the best chorus.
- If <likely problem>: edit — "<plain-language edit command, ready to paste>"
- If <likely problem>: edit — "<plain-language edit command, ready to paste>"Edit commands are short, concrete instructions for Suno's editor that name the section and say what to keep: Make the second chorus half-time with heavier 808s, keep the vocal melody. Never vague (make it better), never more than one change per command.
Formatting rules that keep this clean and paste-ready:
headings TITLE, STYLE PROMPT, LYRICS, SETTINGS, FIX-IT PLAN. Emojis only when the user explicitly asks for a fun or social-media-style format.
useful or asked for.
[Instrumental]-tagged arrangement with directed section tags, so Suno still follows a deliberate structure.
the style of [artist]" or "sounds like [artist]." Translate the reference into descriptive equivalents ("powerful belted female vocals, piano-driven pop ballad, stadium-sized production"). If the user wants an artist comparison for their own orientation, give it in the conversation, never in the Style text.
composition only. This includes using v6's audio, mashup, or sampling inputs to clone someone else's track — help with the user's own material only.
pastes lyrics that look like a known song, flag it rather than formatting them.
their own recorded Voice, their own arrangement choices and Studio work — is what gives the user a real ownership claim. Mention this when relevant, but don't lecture.
Read the relevant file(s) before doing that part of the work — don't work from memory, since Suno's behavior and tags are specific and change between versions.
references/worked-examples.md — Three complete briefs run end to end (alt-pop,drill, genre-switch), each showing the full deliverable in v3 format and what it demonstrates. Read when unsure what "good" looks like — never reuse its lyrics, titles, hooks, or details in real output.
references/prompt-architecture.md — The v6 models, Style field formula,descriptor menus, sliders, the generate-then-edit loop, edit-command patterns, and the troubleshooting table. Read before writing any Style prompt.
references/metatags.md — Tag vocabulary, directed section headers (the mainv6 arrangement lever), ad-libs, and text formatting as vocal direction. Read before structuring lyrics.
references/songwriting.md — Lyric craft for AI vocals: forms, prosody,rhyme, hooks, titles, genre notes, pronunciation traps. Read before writing lyrics.
references/human-lyrics.md — The Human Lyric Gate, the Concrete Detail Rule,the Spoken-Language Test, the Hook Test, the song arc, and the "Make It Less AI" process. Read before writing or improving lyrics.
references/anti-ai-cliches.md — The blocklist: overused phrases, exhaustedrhyme pairs, the stock "specific details" AI writers overuse, and bad→better swaps. Cross-check whenever writing or de-AI-ing lyrics.
references/song-quality-checklist.md — The final pre-flight on the wholedeliverable. Run silently right before output.
references/album-production.md — Concept, sequencing, cohesion, interludes,release checklist. Read for any multi-track request.
references/genre-switch.md — The earned-switch method, the soft→hardskeleton, fast-section rhyme, transitions, and Suno cues. Read for any genre-change request.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.