suno-pack

name

suno-pack

description

This skill should be used when the user wants to create music with Suno (AI music generation) — a song, track, beat, anthem, jingle, or instrumental — or wants to execute an existing suno-pack for real via the suno-pp-cli integration — rendering packs into actual tracks, cover pipelines, Suno library checks, one-roll experiment lanes, and saga-synced field journals. Produces a complete track package — concept document, per-version style and lyrics prompts, instrumental variants, cover workflow, runnable scripts, and an experiments lane book. Supports "--versions 5.5 4.5" and "--mode experimental". Responds to "suno-pack", "make a track in Suno", "write me a song", "Suno prompt", "lyrics for Suno", "instrumental version", "render the pack", "make it real", "generate it for real", "run the cover pipeline", "how are my Suno tracks doing", "suno experiment", "sync the journal", "rebuild the experiment saga", or any request to turn an idea, mood, or theme into Suno-ready prompts or finished Suno tracks.

Suno Pack

Turn a track idea into a complete, paste-ready Suno package: a concept document that gives the track a soul, then per-version prompt artifacts that express it — plus the cover pipeline that combines v4.5's prompt obedience with v5.5's audio quality.

Core principle: concept before prompts. A great track is not a pile of genre tags — it is one emotional idea expressed through sound. Write the concept first; derive every prompt from it. The prompts are the expression of the concept, the same way a canvas expresses a design philosophy.

REQUIRED REFERENCES: Read references/suno-v5-5-prompting.md before writing any prompt — it carries the shared truth: field limits, the canonical metatag taxonomy, per-version behavior and slider settings, the cover workflow, and the instrumental anti-vocal-hallucination kit. When v4.5 is among the requested versions, ALSO read references/suno-v4-5-prompting.md for the v4.5-specific prompt styles and lore. Before authoring vocal lyrics, rewriting lyrics, or diagnosing lyric density/flow problems, read references/lyrics-rhythm-craft.md — it is the lyric beat-map and performance notation contract. Before ANY execution work (rendering, covers, library, experiments), read references/pp-cli.md — the verified command truth; for experiment mode additionally references/experiment-lanes.md. Do not prompt from memory and do not improvise CLI flags — model behavior, limits, lyrics, and the CLI surface are all version-specific.

When to Use

• "Make me a track / song / beat about X"
• "I need Suno prompts for ..." / "lyrics for Suno"
• "Instrumental version of this idea"
• Any creative brief — a mood, a scene, a story, a product, a feeling — that should become music.

Not for: composing sheet music, audio post-production advice, or non-Suno generators (adapt the concept phase, but the prompt syntax is Suno-specific).

Arguments

Parse from the invocation text; accept natural-language equivalents.

Argument	Default	Effect
--versions <v...>	5.5 4.5	which model versions get prompt files (accepts 5.5 5.0 4.5, comma-separated, or "only v5.5", "all versions")
--no-cover	cover on	skip the cover/re-render file
--cover <from> <to>	oldest→newest requested	override the cover chain endpoints
--mode experimental [n|list]	faithful	experiment mode: random lane, lane n, or the lane menu — see references/experiment-lanes.md

Cover file rules: generated whenever ≥2 versions are requested — from the oldest requested version (the authoring model, best prompt adherence) to the newest (the rendering model, best audio). One version requested → no cover file unless --cover says otherwise.

Workflow

1. Absorb the brief

Take whatever the user gives — a phrase, a mood, a story, a genre. Parse arguments. Infer boldly; ask only when a choice genuinely forks the work (e.g. vocal language, explicit genre preference). If the user gave a track name, keep it; otherwise name the track yourself — evocative, short, no genre words.

2. Write the concept (concept.md)

This is the design-document phase — the track's manifesto. 4–6 short sections, written with conviction, generic enough to survive regeneration but specific enough that two readers would imagine the same track:

• Premise — the one-sentence soul of the track. What it is about, emotionally, not musically.
• Mood & emotional arc — where it starts, where it peaks, how it ends. An arc, not a static adjective list.
• Imagery & motifs — the pictures the track should put in the listener's head. Concrete images (sodium streetlights, a coffee going cold), recurring motifs, the one detail that makes it specific.
• Sonic palette — era, instrumentation, textures, production character, tempo feel, and the lead voice (human or instrument) that carries the melody. Adjective+instrument pairs.
• Structure sketch — the intended arrangement arc in plain language (how it opens, where the energy lifts, what contrast the bridge brings, how it closes).
• Generation notes — which versions are being targeted and why, the recommended pipeline (e.g. author on v4.5 → cover on v5.5), how many takes to budget per version, post-production flags (de-esser, mastering), and any niche-genre workaround that applies.

2.5 Write the lyric beat map (vocal/lyrics work only)

For any vocal track, lyric rewrite, or "lyrics for Suno" request, read references/lyrics-rhythm-craft.md before writing the ## Lyrics block. Suno lyrics are not poems; they are a beat map plus performance score.

• Pick a cadence-reference artist by emotional intent before writing. Search for real lyric/form examples when tools are available; do not go from vibes and do not copy lyrics. Extract structure, line lengths, repetition/chant patterns, open-vowel releases, build devices (anaphora, call-and-response, wordless climb), negative space, and delivery/layering ideas, then write original lyrics with that rhythmic DNA.
• After every vocal ## Lyrics block, add ## Lyric Form Inspiration naming the artist(s) and song title(s) used for form/cadence inspiration, plus the specific structural observations borrowed. This is provenance for the human and a test surface for the skill; do not paste or quote lyric lines. If no search/source was available, say so instead of inventing references.
• Build the section skeleton first ([Intro], [Verse 1], [Chorus], [Bridge], [Final Chorus], [Outro], [End]).
• Keep corresponding lines roughly within the ±2 syllable rule. Use short verses, tighter chantable choruses, and open-vowel release lines where the voice should soar.
• Preserve the source motifs, but spend one image per slot. Translate dense concepts, jargon, and stacked metaphors into singable concrete phrases.
• Use [ ] only for structure, arrangement, delivery, and section cues; use ( ) only for sung backing layers, echoes, ad-libs, or harmonies. Never put production instructions in parentheses.
• Add performance notation while writing: sparse backing echoes in ( ), power words in CAPS (1-3 per section), sustained vowels with ..., ~, or repeated letters, and chorus layering cues such as [Chorus | stacked harmonies] when the concept supports them.
• Verify line shape before claiming the lyric is fixed. If the lyric cadence implies a faster pulse, chant percussion, 6/8 feel, or a different BPM, update the style prompt too so music and words agree.

Pure instrumental/no-lyrics files do not use this as lyric authoring: their Lyrics block remains structure tags only, starting with [Instrumental] and ending with [End], with no verses, sung words, ad-libs, or parenthetical backing vocals.

3. Derive the prompts (one vocal + one instrumental file per version)

Every line of every prompt must trace back to the concept. Follow the reference for syntax, limits, and per-version prompt shape — the style prompt is version-specific, never copy-pasted across versions:

• v5.5 / v5.0 — detailed descriptor stack (4–7 categories, 400–800 chars): genre, era, mood, adjective+instrument pairs, stacked vocal descriptors, production texture, BPM.
• v4.5 — emit ALL prompt-style variants, each a translation of the same concept (styles defined in the v4.5 reference). The vocal file carries three — A: lean tags, B: conversational-narrative, C: structured-object — the instrumental file two (A and C, with the vocals: key dropped from C; conversational is excluded there because its prose is vocal-contaminated by nature and its structure narration is redundant when the lyrics block already carries arrangement tags). Mark exactly ONE variant Recommended, chosen from the concept: lean → niche genre adherence; conversational → structure-arc briefs; structured-object → era/artist-flavored briefs. All variants ≤500 chars, genre first; repeat the genre name for stubborn niches.

The vocal lyrics field ports across versions unchanged (same metatag system) and must come from Step 2.5: concrete motifs, sparse density, verified line shape, deliberate performance notation, a chorus that earns repetition, and one turn — a line or image that recontextualizes the song near the end.

Every prompt file carries a Settings block with the per-version slider defaults from the reference (adjusted by the concept's Generation Notes).

4. Write the cover file (default on)

cover_<from>_<to>.md documents the re-render pipeline: pick the best take from the authoring version, Cover it on the rendering version with a minimal style prompt (genre, era, production character, BPM — the audio carries structure and lyrics) and the Audio Influence sweep protocol from the reference. This is the quality path for prompt-obedience maximalists.

5. Emit the package

Create ~/.limitless/suno-pack/suno_<slug>/ by default — <slug> is the track name, lowercased, spaces to underscores, ASCII only (e.g. "Sodium Lights" → suno_sodium_lights/). Write elsewhere only when the user explicitly chooses a path. Default file set:

~/.limitless/suno-pack/suno_<slug>/
├── concept.md
├── lyrics_v5.5.md
├── no_lyrics_v5.5.md
├── lyrics_v4.5.md
├── no_lyrics_v4.5.md
├── cover_4.5_5.5.md
├── experiments.md                 # concept-derived lane book: web-UI
│                                  #  recipes + field journal (template +
│                                  #  derivation rules in
│                                  #  references/experiment-lanes.md)
├── generate_lyrics_v5.5.sh        # one runnable script per prompt file
├── generate_no_lyrics_v5.5.sh     # (template in references/pp-cli.md;
├── generate_lyrics_v4.5.sh        #  keep in sync when prompts change)
└── generate_no_lyrics_v4.5.sh

experiments.md payloads are DERIVED from the concept (read the derivation rules in references/experiment-lanes.md before emitting) — pre-rolled picks with re-roll menus, never generic madlib fills. Executed packs accumulate audio/ and runs/.

Then present a compact summary: track name, premise line, file list, and the output path, recommended pipeline order, ending with the handoff line: the pack is paper until the user says so — "say 'make it real' to render it (credits, I'll confirm cost first), or pick a lane from experiments.md / 'give me experiment N'". Offer variations (different genre lens, language, vocal swap) as a follow-up, don't generate them unasked.

Make It Real — Executing a Pack

Triggers: "render ", "generate the pack", "make it real", "run the cover pipeline", "how are my Suno tracks doing". Read references/pp-cli.md first — it is the only source of CLI truth; never improvise or guess flags. The execution loop:

1. Gates: binary present → doctor healthy → credits checked. Any gate fails → the degradation playbooks in the reference (offer install / browser re-auth; never paste secrets; fall back to authoring-side work — never fabricate execution).
2. Parse the prompt file: Settings table → flags; ## Lyrics fenced block; ## Title; ## Exclude Styles; the ★ Recommended style variant unless the user names one.
3. Already-ran check: pack runs/ hashes + library grep — surface prior takes before spending.
4. Preflight: the exact command with --dry-run (free, validates flags).
5. Confirm: one line — estimated cost, credit balance, running pack total, prior-takes finding. Explicit user yes required for EVERY spending/mutating command; --yes never substitutes for it.
6. Fire and land: run with --wait; parse clip ids; per-clip download + take-aware rename (<slug>-<model>-take<N>-<clipid8>.mp3 — NEVER --download, it collides on title); write the immutable run log to runs/; sync --latest-only.
7. Report: files, clip ids, credits before/after, captcha state. The human listens and judges; the skill never claims to have heard anything.

Cover pipeline ("run the cover pipeline"): same loop — seed picked by the human from skill-proposed observable-cited candidates, cover command truth and the no-slider-flags caveat per the reference, parent_clip in the run log, lineage verified after sync.

Experiment Mode

--mode experimental — read references/experiment-lanes.md plus references/pp-cli.md first. One invocation = one lane = ONE roll; cost stated before firing; unmet requirements (e.g. missing seed) are resolved in conversation, never refused or silently re-rolled. The pack's experiments.md is the menu source and journal ledger (emit it first if the pack predates it). After the roll: run log with lane metadata, journal row appended unverdicted, stop — depth (re-rolls, taming, sweeps) is always human-triggered, one invocation each.

Verdicts are the field journal scale — love/like/nope/hate, keep = like or better; no numeric scores on art, ever. ♥ in the Suno UI syncs as "like" via saga sync.

Saga Sync — Rebuild the Journal from the Library

Triggers: "sync the journal", "sync the pack", "rebuild the saga", "update the journal from my library". Read-only, free, no confirmation needed — recipes in references/pp-cli.md (Saga sync section). The loop: clips by pack title(s) → lineage closure via metadata.cover_clip_id (rebuilds the full tree offline, including rolls made by hand in the web UI) → strays by style/time-window surfaced as QUESTIONS, never silently included or dropped → journal merge under the sacred rules: one row per generation, is_liked → at least "like" (mark which clip carries the ♥), never downgrade, never overwrite a human verdict or note, absence of a like is never "nope" → render the lineage tree, update the running credit total.

Artifact Specs

Field order in every prompt file mirrors Suno's Custom Mode UI top-to-bottom, so the user can paste while scrolling: Settings → Lyrics → Lyric Form Inspiration (read-only provenance; do not paste to Suno) → Style → Exclude → Title.

concept.md

# <Track Name>

*Track concept — <date> — suno-pack*

## Premise
## Mood & Emotional Arc
## Imagery & Motifs
## Sonic Palette
## Structure Sketch
## Generation Notes

lyrics_v<version>.md — vocal version

# <Track Name> — Vocal — v<version>

## Settings
| Setting | Value |
|---|---|
| Model | v<version> |
| Mode | Custom |
| Instrumental toggle | OFF |
| Weirdness | <n>% |
| Style Influence | <n>% |
| Takes to budget | <n> |

## Lyrics
```text
[Intro: <texture descriptors>]

[Verse 1]
...

[Chorus]
...

[Bridge: <contrast descriptors>]
...

[Outro: <wind-down descriptors>]
[End]

Lyric Form Inspiration

Artist	Song(s) studied	What was studied	How it shaped these lyrics
	<song title(s)>	<line length, repetition, open vowels, build, negative space, layering>	<specific original-form decision; no copied lyric lines>

Style of Music

<version-appropriate style prompt — detailed stack for v5.x,
lean genre-first variant for v4.5>

Exclude Styles (Advanced Options)

<comma-separated unwanted genres/elements>

Title

<track title, ≤80 chars>

**v4.5 deviation — style variants:** in `lyrics_v4.5.md` the
`## Style of Music` section carries three labeled variants instead of one
block (two in `no_lyrics_v4.5.md`: A and C only). Shape:

```markdown
## Style of Music — pick one variant, test in order

### A — lean tags · wins on niche genre adherence
```text
<lean variant>

B — conversational · wins on structure-arc briefs · ★ Recommended

<narrative variant>

C — structured-object · wins on era/artist flavor

(reset Advanced Options before pasting)

<const-block variant>

Testing discipline: one roll per variant, judge on pairs (a flaw must repeat to count), then iterate only the winner.

The ★ Recommended marker goes on whichever variant the concept favors —
users who don't want to experiment paste that one and ignore the rest.

### `no_lyrics_v<version>.md` — instrumental version

Same layout and field order, with the full anti-vocal-hallucination kit from
the reference:

- Settings block is SURFACE-AWARE (web UI toggle and Lyrics box are
  mutually exclusive; v5.5 UI has no toggle — see the anti-vocal kit in
  the v5.5 reference): two rows —
  `Instrumental — web UI | v4.5: toggle OFF, paste the Lyrics block (ON disables the box) · v5.5: no toggle, the Lyrics block is the defense`
  and `Instrumental — CLI | --instrumental flag + Lyrics block (both transmit)`.
- Lyrics block contains structure tags only — canonical instrumental and
  dynamics tags, opening with `[Instrumental]`, closing with `[End]`.
- Style block rewritten with zero vocal-adjacent words; the concept's melody
  carrier becomes an instrument description (version-appropriate shape).
- Exclude block includes `vocals, singing, choir, spoken word, ad-libs`.

### `cover_<from>_<to>.md` — re-render pipeline

```markdown
# <Track Name> — Cover Pipeline v<from> → v<to>

Author on v<from> (prompt obedience), render on v<to> (audio quality).

## Pipeline
1. Generate on v<from> using lyrics_v<from>.md / no_lyrics_v<from>.md
   until a take nails song, structure, and feel (budget <n> takes).
2. Open the winning take → Cover → switch model to v<to>.
3. Paste the minimal style prompt below. Do NOT re-enter lyrics — the
   audio carries them.
4. Sweep Audio Influence per the protocol below.

## Settings
| Setting | Value |
|---|---|
| Model | v<to> (Cover) |
| Weirdness | 10–20% |
| Style Influence | 20–40% |
| Audio Influence | start 50% — sweep, see below |

## Cover Style Prompt (minimal — genre, era, production, BPM only)
```text
<minimal re-render prompt>

Audio Influence Sweep

Cover behavior shifts between Suno backend updates — treat the slider as a dial to calibrate, not a constant. One roll at 50%, then:

• output hugs the source too hard (near-clone) → drop toward 30%
• output drifts away from the song → raise toward 80% Change nothing else while sweeping. Expect 2–4 rolls.

## Quick Rules

| Rule | Why |
|---|---|
| Style prompt is per-version: detailed for v5.x, variant menu for v4.5 | the models obey different prompt shapes |
| v4.5 files ship all style variants with one ★ Recommended | which dialect wins is empirical; one roll each, iterate the winner |
| Style ≤1000 chars (sweet spot 400–800 on v5.x, ≤500 on v4.5), genre first | overflow silently truncated; unrelated tag piles average into mush |
| Canonical metatags only; `[Tag: descriptors]` for per-section direction | invented tags get ignored or sung |
| Performance directions in bracket tags, never bare `(parentheses)` lines | parens are sung as backing vocals |
| Negations via Exclude field + affirmative style phrasing | v5.x largely ignores "no X" in style |
| No artist names, no mixing jargon, BPM as number | unreliable / ignored / approximate |
| Instrumental = toggle + `[Instrumental]` + clean style + exclude vocals | v5.5 hallucinates vocals; layer all defenses |
| Repeat identical chorus text; number the verses | melodic reinforcement, structure tracking |
| `[End]` on everything | prevents trailing audio |
| Cover style prompt stays minimal; sweep Audio Influence | style field weakly honored in covers; slider semantics unstable |
| Execution: explicit user yes before EVERY spend, cost stated first | `--yes` silences the CLI's prompt, not the user's |
| Never `--download` on generate; per-clip download + take-aware rename | title-named files collide — a paid take was destroyed this way (2026-06-11) |

## Common Mistakes

- **Prompting from memory of older Suno versions** — limits and behavior
  changed; the reference is the truth.
- **Skipping the concept** and stacking genre tags — produces competent,
  soulless output. The concept is where "amazing" comes from.
- **Copy-pasting the v5.5 style prompt into the v4.5 file** — v4.5 wants
  lean and genre-first; the detailed stack degrades adherence there.
- **Reusing the vocal style prompt for the instrumental** with "no vocals"
  appended — v5.5 will sing anyway. Rewrite it clean.
- **Stage directions as parentheticals** (`(whispered)` on its own line) —
  use `[Bridge: whispered, stripped down]`.
- **Restating lyrics or structure in the cover prompt** — the source audio
  carries both; a long cover prompt only adds drift.
- **Improvising pp-cli flags** (`--style`, `--prompt`, `--exclude-styles`)
  or probing generate commands with `--help` — the flag truth lives in
  `references/pp-cli.md`; validate with `--dry-run`, never a live call.