SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillssuno

suno

Agents & models

View repository

Updated 10-16 00:24

About this skill

Variables

Arguments: $ARGUMENTS

Purpose

Suno is an AI music platform; the prompt is the instrument. This skill helps the user generate, refine, and debug Suno prompts — both the style/genre and lyrics fields — using v5.5-era best practices, the documented tag taxonomy, and community-validated performance tricks.

The skill is a prompt-craft assistant, not an audio generator. Suno produces audio; this skill makes sure the prompt going INTO Suno is sharp, on-format, free of common pitfalls.

Suno v5.5 (released March 26, 2026; verified 2026-07-18 against https://suno.com/blog/v5-5) preserved v5's prompt syntax but improved adherence to nuanced descriptors and added three personalization layers (Voices, Custom Models, My Taste). Everything below targets v5.5 unless noted; legacy v4 (~200-char style prompt) is out of scope.

Action Router

Parse $ARGUMENTS: first token = action, remainder = args. If empty (e.g. model-invoked from a natural-language request), infer the action from conversation context; show this menu only when no actionable context exists.

Action Purpose Detail
prompt <intent> Build a complete style prompt + lyrics from a song idea inline + context/style.md + context/lyrics.md
lyrics <intent> Generate or refine lyrics with section tags, vocal tags, per-section style overrides, performance cues context/lyrics.md
style <intent> Generate a style prompt using the 6-layer formula context/style.md
clean <text> Review user's existing prompt, flag pitfalls, propose rewrite context/troubleshoot.md + context/style.md
research <topic> On-the-fly external research — artist sonic profiles, current trends, niche genres, specific reference songs (BPM/key/instrumentation). Translates findings into Suno descriptors. context/research-recipes.md
tags Show the full tag taxonomy (structural, vocal, instrument, mood, ending) context/lyrics.md
genre <name> Genre handler — 4 modes: load template (12 built-ins), research a niche genre, suggest genres for a vibe, or show a genre family tree templates/.md + context/genre-taxonomy.md
troubleshoot <symptom> Diagnose a specific failure (hallucinated lyrics, BPM ignored, vocal artifacts, exclusions failing) context/troubleshoot.md
tips Community-validated lyric-side performance tricks (caps, vowel stretch, ellipses, parentheticals) — MEDIUM confidence context/tips.md
power-tips Empirically-validated power-user techniques (tag weighting, genre fusion order, stem-loop refinement, two-pass vocal isolation, persona-as-draft, punctuation cadence, describe-this-song, Custom Model tricks) — MEDIUM-HIGH confidence context/power-tips.md
features v5.5 feature index — Voices, Custom Models, Personas, Cover, Extend, Replace, Stems, Studio, Sliders context/v55-features.md + context/advanced.md
voices Voice cloning deep dive — recording, verification, conflict rules with style prompt context/voices.md
studio Suno Studio (GAW) — what's possible, clip ops, Take Lanes, Warp Markers, MIDI, stems context/studio.md
workflow <recipe> Demo upload, edit/rearrange, add instruments, vocal comping, multi-genre cover chain — 8 recipes context/workflow-recipes.md

If the action is unknown, show this table and ask what the user wants.

Load-bearing fundamentals (inline — every action assumes these)

The 6-layer formula (style prompt ordering)

Order matters — early tags carry more weight. Front-load genre.

  1. Genre / subgenre — specific, not "pop" → synth-pop, 80s-inspired
  2. Mood — 2-3 words → nostalgic, hopeful (not 9 conflicting moods)
  3. Instrumentation — specific → fingerpicked acoustic guitar (not just "guitar")
  4. Vocal direction — acoustic descriptors → breathy, intimate, slight rasp (not "amazing")
  5. BPM — numeric → 128 BPM (not "fast")
  6. Production — polished mix, lo-fi tape hiss, compressed and aggressive

Sweet spot: 5–8 distinct tags. Fewer than 4 → generic. More than 10 → conflicting signals.

Character budgets

Field Limit Notes
Style prompt (v5/v5.5) ~1,000 chars Up from ~200 in v4. Truncates silently. Front-load critical content.
Lyrics 5,000-char hard cap (v4.5/v5/v5.5) 3,000 was the v4-era cap. Quality sweet spot stays ~3,000 (~40-60 lines / 200-300 words) — past that Suno rushes, skips sections, or cuts output short.
Title ~100 chars Up from ~80 in v4. No effect on musical output
Exclude field Free-text in Advanced Options (Custom mode) Same vocabulary as inline negatives

Verified 2026-07-18 — no official Suno page states field limits; figures are third-party tester consensus (hookgenius character limits, aimusicapi cheat sheet, 2026-07-03).

The lyrics field is a SECOND style channel

Power users treat the lyrics field as more than words to sing. Per-section style overrides in lyrics control instrumentation, dynamics, and production section-by-section — something the global style prompt can't do.

[Verse: piano only, no drums, intimate]
First narrative beat...
The vocals carry the weight...

[Pre-Chorus: drums enter, building]
Tension rising...
Tension rising...

[Chorus: full band kicks in, distorted guitar lead, gang vocals]
THE HOOK
THE HOOK

[Bridge: stripped to vocals + acoustic, dry mix]
Most vulnerable moment...

[Chorus: bigger than before, double-tracked vocals, layered harmonies]
THE HOOK

Two separator syntaxes work — both on the same [Tag] line:

  • [Verse: descriptor, descriptor] — colon + comma list
  • [Bridge | dark | introspective | sparse] — pipe-separated

Keep modifiers short (2-4 words each). Match section to tone — "explosive drop" only makes sense on [Drop], not [Intro].

This solves the dynamics problem — without it, every section sounds the same density. With it, you can build verse-chorus contrast, strip the bridge, blow up the final chorus.

Worked example (synth-pop)

Style prompt:

synth-pop, 80s-inspired, nostalgic and hopeful,
shimmering analog synths, warm Moog bass, punchy drum machine,
ethereal female vocals with reverb tail, 95 BPM,
polished radio-ready production, no autotune

Lyrics (skeleton):

[Intro]
[Verse 1]
First narrative beat, present tense, concrete imagery...

[Pre-Chorus]
Building tension, single-line repetition encouraged

[Chorus]
The HOOK — peak energy, full instrumentation
Repeat the title or core phrase

[Verse 2]
Second narrative beat, contrast or escalation

[Chorus]

[Bridge]
Contrast section — often a key change or stripped-back

[Chorus]
[Outro]
[End]

Negative prompts go at the end of the style prompt: ..., no autotune, no reverb wash. The Exclude field in Custom mode's Advanced Options is the structural alternative.

Anti-patterns (always flag these in clean action)

Anti-pattern Why it fails Fix
Vague genre (pop, rock) Generic AI sound Specify subgenre + era + tonal cue
5+ stacked genres Muddy mix, incoherent style Pick 1-2 max, or use a clear hybrid (nu-metal dubstep)
9+ mood words Conflicting emotional signals 2-3 related moods
amazing, epic, beautiful Zero effect on output Replace with acoustic descriptors (raspy, breathy, intimate)
BPM as descriptor (fast) ±20 BPM drift Use numeric (128 BPM) — ~90% adherence in v5.5
(x2) after a lyric line Repeat is largely ignored Write the line twice with minor variation
Naming artists directly (like Drake) Blocked or ignored Use sound descriptors (Toronto trap bounce, silk-smooth R&B falsetto)
no drums in a drum-heavy genre alone Drums still appear Pair with positive (piano only, no drums)
Style prompt > 1000 chars Silent truncation Front-load critical content
Capitalize EVERY word Effect dilutes Cap only turning-point words for emphasis
Vocal descriptor + active Voices/Custom Model Conflict between clone and prompt Drop gender/tone descriptors when a Voice or Custom Model is selected

Custom mode is required for tag syntax

Simple mode = unified prompt + auto-lyrics. Custom mode = separate fields (style, lyrics, title) + Advanced Options + tag syntax + Creative Sliders. Any time the user wants [Verse]/[Chorus]/[Female Vocal]-style tags, they must be in Custom mode. See context/advanced.md.

How to handle each action

prompt <intent> — build both style and lyrics. Read context/style.md for layer detail, context/lyrics.md for tag taxonomy. Produce two fenced blocks (style, lyrics), each labeled, within budget. End with 2-3 tweak suggestions.

lyrics <intent> — load context/lyrics.md. Output a tagged lyrics block with section structure ([Verse 1], [Chorus], etc.). Include 1-2 performance cues (parentheticals, ellipses) where they fit naturally. Stay ≤3,000 chars — the quality sweet spot; the hard cap is 5,000 — unless user asked for max.

style <intent> — load context/style.md. Walk the 6-layer formula. Output one fenced style-prompt block + character count. Suggest 2-3 alternatives (different mood, different production).

clean <text> — load context/troubleshoot.md and context/style.md. Score the user's prompt against the anti-pattern table above. Surface flagged issues with specific quotes. Propose a rewrite. Do NOT silently rewrite — show the diff in reasoning.

tags — load context/lyrics.md. Render the full taxonomy as tables (structural, vocal, instrument, mood, ending). Include parameterized syntax [Tag: descriptors] example.

genre <name> — 4-mode router. Detect intent from arguments:

  1. Template mode — <name> matches one of the 12 built-in templates (pop, rock, hip-hop, trap, edm, jazz, classical, folk, metal, ambient, lofi, rnb): read templates/.md and present.

  2. Research mode — <name> is a real genre but no template exists (e.g., dungeon synth, witch house, gqom, phonk, slowcore, vaporwave, nu-disco):

    • Check context/genre-taxonomy.md first — if cataloged, build a Suno prompt from the catalog entry
    • If not cataloged, invoke the research action's Phase 1 (WebFetch / Perplexity) to fetch BPM range, instrumentation, vocal style, production character
    • Synthesize into the standard template format: style prompt block + lyrics shell + 3 tweak knobs + 2-3 common variants
    • Cite sources

  3. Suggest mode — <name> starts with suggest (e.g., genre suggest dark moody chill, genre suggest dystopian aggressive) OR is a vibe phrase rather than a genre name:

    • Read context/genre-taxonomy.md "Vibe → genre mapping" section
    • Return 3-7 fitting genres, each with: 1-line distinguishing characteristic + the right BPM/key/instrumentation hints + which template to load OR which research target to chase
    • Ask user which one to expand into a full prompt

  4. Family mode — <name> starts with family (e.g., genre family hip-hop, genre family electronic):

    • Read context/genre-taxonomy.md family tree
    • Show the requested family's subgenres in tree form
    • Brief 1-line characteristic per subgenre
    • Ask user which leaf to expand

If <name> is ambiguous (could be template OR research OR suggest), ask user which mode they want.

troubleshoot <symptom> — load context/troubleshoot.md. Match the symptom to a known failure mode. Give diagnosis + 1-3 specific fixes + a regenerate-with-this prompt rewrite if applicable.

tips — load context/tips.md. Surface community-validated techniques (caps for vocal pressure, vowel stretching for melisma, ellipses for breath, parentheticals for inline directives, hyphenation for staccato, line breaks for melodic separation). Caveat: MEDIUM-confidence — consensus across multiple community sources but no official Suno doc.

features — load context/v55-features.md (what's new in v5.5) and/or context/advanced.md (modes, Personas, Cover, Extend, Replace Section, Stems, Sliders, Studio). Pick the section the user asked about — don't dump everything.

Confidence flags (be honest about source quality)

  • HIGH confidence: claims confirmed by Suno's official help center (help.suno.com) or suno.com/blog. The 6-layer formula, structural and vocal tag names, Custom mode requirements, Voices/Custom Models/My Taste mechanics, Creative Slider behavior — all HIGH. Character budgets are NOT officially published — third-party tester consensus only (MEDIUM-HIGH).
  • MEDIUM confidence: community-validated techniques across multiple guides + Reddit consensus, but no official Suno doc. Capitalization weighting magnitude, vowel-stretching letter counts, hyphenation staccato, (x2) failure mode — all MEDIUM. Effects are real (multi-source agreement); exact magnitudes are folk wisdom.
  • Re-verified 2026-07-18 — position flipped since the 2026-05-10 pass: current third-party testers agree the lyrics hard cap is 5,000 chars on v4.5/v5/v5.5; 3,000 was the v4-and-earlier cap. The earlier "3,000 consensus" conflated the old cap with the quality threshold. ~3,000 remains the practical budget — past it Suno rushes, skips sections, or shortens output. Sources: hookgenius character limits, aimusicapi cheat sheet, 2026-07-03. No official Suno page states field limits.

When generating prompts, default to HIGH-confidence techniques. Surface MEDIUM-confidence tricks as opt-in suggestions, not commands.

What this skill does NOT do

  • Generate audio — Suno does that. This skill produces text prompts.
  • Cover legacy v4 prompting — the ~200-char style-prompt era is out of scope. Targets v5.5.
  • Make commercial-use determinations — Cover and Voices have commercial constraints noted in context/advanced.md; treat as informational, not legal advice.
  • Replace Suno's web UI workflow — this is a prompt-craft assistant. Workflow steps requiring clicks in Suno's UI (Replace Section, Persona creation, Voice verification) are described, not automated.
  • Fix already-generated audio — Suno's prompt is a generation-time control. Most issues (hallucinated lyrics, vocal artifacts, BPM drift) are prevention-only. The skill's troubleshoot action gives the prompt-side fix, then asks the user to regenerate.