npx skills add ...
npx skills add minimax-ai/minimax-music3 --skill music-caption-rewriter
Turn a brief music description and optional tagged lyrics into a professional MiniMax Music 3 structured caption with Global Metadata, Vocal Details, and a section-aware Arrangement. Use when users ask to enhance a music-generation prompt, preserve lyric-section directives, retrieve a similar style from bundled templates, fuse styles, or produce JSON or JSONL caption output.
npx skills add minimax-ai/minimax-music3 --skill music-caption-rewriter
Transform the user's musical intent into a new, generation-oriented structured caption. Find useful references through progressive disclosure: route to a small style family, compare compact cards, then read only the selected complete templates.
Use natural-language reasoning and local text files only. Do not execute scripts, build a database, calculate embeddings, call external APIs, or scan all 1,000 templates.
Accept:
Caption: required natural-language music description.Lyrics: optional lyrics containing bracketed section or control tags.Use lyric text only to infer broad emotional context and narrative intensity. Never quote, paraphrase, summarize, or reproduce it. Treat only bracketed tags as executable structural, musical, vocal, or production directives.
Follow these stages in order:
Do not expose the Music Brief, routing choices, scores, or template IDs unless the user requests diagnostics.
Extract only supported or reasonably inferred values:
Classify each value internally as explicit, tagged, inferred, or unspecified.
Do not invent a precise key, BPM, vocal gender, melodic interval, or production technique when a broader description is sufficient.
Preserve an explicit instrumental request. Do not add vocals. If vocal presence is unspecified, choose a conservative treatment supported by the user's description and the closest style family.
Apply this precedence:
A section tag may change its local arrangement without replacing the song's global genre. Preserve a hard user exclusion when a tag conflicts with it.
When two explicit instructions conflict, prefer the more specific and later instruction if the intent remains clear. Otherwise make the smallest musically coherent compromise.
Never silently reverse an explicit vocal gender, instrumental requirement, tempo limit, required instrument, or prohibited element.
Read the genre router first. Choose:
Use genre, groove, instrumentation, and cultural context as stronger routing signals than generic adjectives such as emotional, epic, dark, or modern.
Read only the family indexes selected by the router. Do not inspect every family index, reconstruct a global catalog, or scan every template filename.
Compare cards in the selected family indexes using this priority:
Apply a strong penalty to direct conflicts. Prefer a close musical family over a card that merely shares mood vocabulary.
Select up to three references with different responsibilities:
Foundation: closest overall identity, groove, and songwriting language.Modifier: best source for a requested secondary genre, vocal character, cultural color, or production texture.Arrangement: best source for section development, energy contour, transitions, and instrument lifecycle.Use one or two references when the request is simple. Do not select a weak match merely to reach three.
Use the Foundation for broad musical identity, the Modifier only for its matched dimension, and the Arrangement reference only for timeline logic.
Do not inherit unsupported details such as a template's exact key, BPM, vocalist, instruments, emotional story, or section order.
Do not copy sentences, distinctive phrases, or a template's complete structure. Synthesize a new caption around the user's brief.
Build around the user's section tags when present. Otherwise choose only sections appropriate to the style, for example:
Intro → Verse → Pre-Chorus → Chorus → Verse → Chorus → Bridge → Final Chorus → Outro
For every included section, state what enters, exits, changes, or intensifies. Keep instrument behavior continuous and make transitions musically plausible.
Create a readable energy arc rather than a static equipment list or a stack of production terminology.
Write the final caption in English unless the user explicitly requests another language.
Return exactly these three top-level headings in this order:
Include genre and subgenres, tempo, emotional progression, and overall sonic and production profile. Use an exact BPM only when explicit or strongly justified; otherwise use a range or qualitative tempo. Include key and scale only when explicit or musically useful.
For vocal music, describe the lead configuration, timbre, register, delivery, harmony or backing vocals, and restrained vocal effects.
For instrumental music, state that the piece is instrumental and identify the instrument or texture carrying the lead melodic role.
Do not invent lyrical subject matter or reproduce lyrics.
Describe the song as a section-by-section timeline. Explain primary and secondary instrument lifecycles, groove development, transitions, embellishments, texture, and spatial effects only where relevant.
Prefer concrete musical changes over decorative prose. Default to approximately 250–450 English words unless the user requests another length.
Do not include a song title, track ID, selected template ID, reasoning trace, or copied lyric line.
Return JSON or JSONL only when explicitly requested. Include original inputs and rewritten_caption. Include routing diagnostics or selected template IDs only when explicitly requested.
Never include complete template contents in machine-readable output unless the user specifically asks for them.
Verify that:
Revise once when any check fails, then return only the corrected result.
Keep the library entirely text-based. When adding a template:
templates/.Do not add scripts, generated catalogs, embeddings, vector stores, databases, or external service configuration.