Recoup Song

Create a song with Recoup

Work in four stages: idea → genre → lyrics and prompt → Recoup API call. The user approves each stage before the next begins. Start at the earliest stage whose current version has not been approved.

Foundation and precedence

Model-specific foundation: recoup-minimax-music-3, including its past-song prompt library. Read references/production-lessons.md for the relevant local findings and their provenance. Use these lessons to improve the draft, not to assume the user's musical choices.

This workflow replaces the foundation's direct-provider call and free re-roll advice: use the Recoup API only, and never generate or re-roll without approval for that call. A song idea is required; a finished video story or paper style is not. If the user wants the music to determine the visuals, leave the visuals undecided until they approve the audio.

Approval contract

• Present only the current stage's concrete proposal and a concise approval question. Stop and wait for the user's reply before doing the next stage's creative work.
• Explicit selection counts as approval of the selected proposal. Preserve prior approvals for unchanged material; don't ask again merely because the conversation resumed.
• A broad request to "make a song," silence, an assistant's own recommendation, or approval of a different stage is not approval of the current stage. A requested revision is not approval of the resulting revision: show it and wait.
• Keep a compact record of approved idea, genre, lyrics/prompt version, and generation scope. An upstream change invalidates approvals for the downstream material it changes.
• Gate 3 approves the exact creative inputs. Gate 4 separately approves spending credits on the reviewed request. Approval for one take never authorizes additional takes.
• Say that these pauses implement the user's requested four-stage workflow. If authorization is ambiguous, keep the action pending.

1. The idea

Ask for the idea if none was supplied, or propose one from the current brief. Present: what the song is about, narrator or point of view, intended listener, emotional change, and the central hook idea. Include the intended use and language. Let the idea develop without proposing a time limit; record a creative length constraint only when the user supplies one.

Do not silently reuse a previous song concept. Do not pick a genre or write lyrics yet. Ask: "Is this the song idea you want to develop, or what should change?" Wait for approval.

2. Genre — study, recommend, then let the user choose

Read the coverage table and relevant production lessons below before recommending a genre. Show two or three plausible choices, one recommendation, and why each serves the approved idea. Include tempo/groove, vocal or instrumental character, and core instrumentation as proposed attributes. Respect an explicit user genre over template density.

"Top" here means most represented in the official caption library, not a measured ranking of generation quality, training-data proportions, or commercial popularity. No genre-quality leaderboard was established by the sources reviewed. Separate local listening results from coverage. Use fit to the idea first, then local evidence and useful template support. Density is only a tie-breaker.

MiniMax Music 3 genre coverage

Verified 2026-09-07: 1,000 cards across 18 families, counted from the official family indexes at commit 945655064d59b98004dd70002e7eb5c8c6e11373. Every count matches its index header. Source links and counting method are in references/genre-sources.md. Counts are a dated snapshot; refresh before claiming they describe a newer library. The older Recoup note overstates the listed family counts by two; this table uses the verified indexes.

For prompt development, consult the official genre router linked in references/genre-sources.md. Choose one primary family, with a secondary only for an intentional blend. Mood words alone are not genre evidence. Prefer the user's identity and groove over generic "pop" routing.

Ask: "Which genre direction should we use?" Wait for their choice or approval. Do not draft the song's lyrics or full generation prompt before this gate passes.

3. Lyrics and prompt

Use the approved idea and genre. Draft original lyrics in the approved language with clear, singable phrasing. Use section tags on their own lines, normally [Intro], [Verse], [Pre-Chorus], [Chorus], [Post-Chorus], [Bridge], [Instrumental], [Solo], [Outro]. Keep production prose in the prompt, not in lines that could be sung.

Let the lyrics and arrangement determine the song's natural length. Reconcile syllable load, phrase lengths, section count, tempo, and instrumental gaps without choosing a shorter API duration from an estimated runtime. The house estimate of ten sung lines per minute is only a starting heuristic, not a model constraint. Keep repeated choruses stable unless variation is intentional, and write a complete ending without an extended empty outro.

Default to the maximum supported API duration: currently duration: 300 seconds. Send it explicitly; omitting the field uses the API's shorter 60-second default. This gives the model room to finish the full song instead of requesting an estimated 45-, 60-, or 150-second runtime that may cut it off. It is generation headroom, not a requirement to compose five minutes of music: do not pad lyrics, repeat sections, stretch the outro, or instruct the prompt to fill 300 seconds. Request a natural ending after the approved sections. Use a shorter API duration only when the user explicitly requests that limit. A short video or an estimated song length alone does not override this default. Recheck the supported maximum before generation; maximum headroom does not guarantee a complete take.

Build the prompt using the official music-caption-rewriter method: read the primary family index, optionally one secondary, then at most three compatible full templates linked from those indexes. Synthesize around this brief; do not copy template sentences or inherit an unrelated vocalist, exact key, BPM, or emotional story. If source access is unavailable, disclose the limitation and use the structure below; never claim template research occurred.

The exact prompt contains three sections:

• Global Metadata: approved genre, tempo/groove, meter and key when chosen, emotional arc, sonic character and production approach. Label deliberate proposals rather than presenting unchosen BPM/key as user requirements.
• Vocal Details: voice, register, delivery, harmonies and effects. Honor an instrumental request and identify the lead instrument instead of adding a vocalist.
• Arrangement: a timeline matching the lyric sections, with concrete instrument entrances, exits, energy changes, transitions and ending. Keep lyric lines out of the prompt.

Show the entire lyrics and exact prompt, plus the API duration (default: maximum generation headroom, currently 300 seconds) and any optional seed. Distinguish API headroom from any estimated musical runtime. Ask: "Do you approve these exact lyrics, prompt and API duration, or what should change?" Wait. Do not run the API or rewrite an approved prompt silently.

4. Recoup API call

Read references/recoup-api.md. Resolve authentication and the intended account using read-only checks. Prepare the exact request with the approved prompt, lyrics and duration. Present:

• POST https://api.recoupable.dev/api/music and a sanitized request preview;
• the resolved account being charged, requested duration, and number of takes (default one);
• the current Recoup credit precheck/maximum quote for the requested duration and its source, or explicitly state if it cannot be verified. Explain that successful billing uses actual generated duration, capped at the quote; if the provider reports no usable duration, the API falls back to the requested-duration charge. Do not promise actual-only billing without that qualification or substitute historical provider cost for the user's Recoup charge.

Ask: "Approve this Recoup API request for one take on this account?" Wait for explicit approval of the displayed request and spending scope. Only then submit, record the returned generation ID, and poll that same generation. Read-only polling is part of the approved call.

Do not re-POST because a request is slow, the network response was lost, or a take sounds wrong. Reconcile pending requests first. A changed payload, a new take, or a different charged account needs new approval for the affected material and call.

After completion, return the playable audio and generation ID, requested versus measured duration, and reported cost when available. Measure the file and compare a timestamped transcript against the approved lyrics when audio tooling is available; disclose anything not inspected. Report omissions, repeats, pronunciation errors, cutoff or genre mismatch instead of calling an unreviewed take finished. Ask the user to listen before using it for a video. Stop at the audio; do not choose the paper style, storyboard, render a video, or publish without a subsequent request.