1The two inputs

Every generation is driven by exactly two texts. They play different roles — keep them separate.

2Lyrics and section tags

[intro]
[verse]
Morning light filtering through the pine
Every quiet street is yours and mine
[pre-chorus]
Hold your breath, the sky is opening
[chorus]
We are made of sound and time
[outro]

• Every tag goes on its own line. Text on the same line as a leading tag is dropped by the model's input contract — [verse] Morning light… silently loses the words.
• A tag may change its section's local arrangement (e.g. [solo]) without replacing the song's global genre.
• Structure is guidance, not a guarantee: tags steer the layout, but the model composes freely within it.
• The requested duration is an upper bound — the model ends the song naturally with a stop token when the lyrics run out. Short lyrics make short songs; a full five-minute song wants a full song's worth of sections.

3The structured caption

For precise control, write the caption under exactly three headings, in this order. Aim for roughly 250–450 words total: specific enough to guide generation, short of an essay.

Global Metadata

• Basic attributes — genre and subgenres, tempo, and (only when you actually want them) key and scale: "bpm is 122. key is G, and scale is minor. Disco / Funk Pop."
• Global emotional progression — the song's emotional arc from open to close, as a story: where it starts, where it peaks, how it resolves.
• Application scenarios & imagery — a scene the song belongs to (a night drive, a dim room, a rooftop countdown). This anchors mood better than adjectives.
• Sonics & production profile — mix character: stereo width, frequency balance, dynamics (e.g. polished and compressed vs. natural and uncompressed).

Vocal Details

• Vocal gender & timbre — always state it explicitly ("Singer A (Female), a warm mezzo-soprano with a breathy low register"). Leaving vocals unspecified is the #1 cause of unwanted instrumental drift.
• Vocal style — delivery and dynamics per section: soft and close-miked in the verse, belted in the chorus.
• Harmony / backing vocals — doubles, stacked harmonies, call-and-response, and where they appear.
• Vocal FX — restrained: reverb, delay throws, saturation, where each applies.
• For instrumental music, say so explicitly and name the instrument carrying the lead melodic role.

Arrangement

• Instrument lifecycle (primary / secondary) — what anchors the song from start to finish, and what enters, exits, or transforms along the way.
• Groove & foundation progression — the rhythmic foundation and how its density evolves section by section: what the drums do in the verse, what lands in the chorus, what strips back in the bridge.
• Embellishments, textures & spatial FX — risers, sweeps, ear candy, reverb tails — only where relevant.

4Writing rules that matter

• Don't fabricate precision. Use an exact BPM or key only when you truly want it; otherwise a range or a qualitative tempo ("driving", "unhurried") gives the model room to be musical.
• Never contradict yourself silently. An explicit vocal gender, instrumental requirement, tempo limit, required instrument, or exclusion should survive the whole caption.
• Precedence when things conflict: explicit requirements first, then section-local tag directives (within their section), then implications of your description, then genre defaults.
• Write an energy arc. A caption that reads as a story — tension, release, respite, climax — generates better structure than a static description.
• Keep lyric content out of the caption. The caption describes music, not words; the lyrics input carries the words.

5Style families

The official prompt library organizes ~1,000 reference captions into 18 style families — each linked below. These are not the limits of what the model can do: they are pre-set references, a vocabulary map for naming genres the way the library does. Any style you can describe is fair game. Words like emotional, epic, dark, modern, cinematic are modifiers — not genres.

Fusing two styles? Name both explicitly ("cinematic orchestral with Chinese folk instrumentation") — the model handles fusions well when the caption commits to how the two palettes share the arrangement.

6Complete examples

Four reference captions from the official library, verbatim. High-energy and intimate, western and East Asian, band-driven and orchestral — all following the same three-heading skeleton at roughly 350 words.

7Pre-flight checklist

• Every section tag sits alone on its own line
• Vocal gender and timbre stated explicitly (or "instrumental" declared, with the lead instrument named)
• Caption has exactly three headings: Global Metadata → Vocal Details → Arrangement
• The arrangement reads as a timeline: entrances, exits, changes, intensities per section
• BPM/key present only if genuinely wanted; no invented precision
• No lyric lines quoted inside the caption
• Explicit exclusions and requirements survive the whole caption without contradiction
• ~250–450 words — specific, not an essay