AI Music Generation and Songwriting

Music creation involves a complex workflow from concept to final product, often hindered for average users due to lack of specialized tools or skills. This skill uses AI to transform themes, lyric fragments, or reference audio into executable creative directions and generates works like songs, instrumental tracks, or BGM, covering scenarios such as ad jingles and multilingual production.

Specific Problems Addressed

When users have musical ideas but cannot independently produce them, such as:
- Needing a structurally complete song generated from themes or lyric fragments.
- Creating new arrangements or style covers based on reference audio (e.g., recordings).
- Generating functional music like background scores, ad melodies, or looped BGM, supporting multiple languages.

How the Skill Works

The skill relies on the Beatra toolchain, executing commands via an MCP client, with key steps as follows:

1. Determining Creative Direction: First, distinguish task types, such as ai-song-cover-studio (covers) or short-video-bgm-studio (instrumental BGM), with this package defaulting to original songs. Reuse confirmed information from the dialogue, and only ask questions when critical choices (e.g., vocal vs. instrumental) are missing.

2. Preparing a Production Card: Before generation, create a concise production card containing purpose, audience, title, emotional arc, genre, tempo, instruments, structure, etc. For example, if genre is empty, ask first, then write the card. Lyrics creation or model queries incur no cost.

3. Selecting and Validating Models: Use model: "suno-5.5" by default, never omit or silently switch to auto. Call beatra.models.list to choose models based on capabilities (e.g., text_to_music), and only use control items returned by that model series—e.g., Suno and MiniMax options are not interchangeable.

4. Using Reference Audio: For local files (e.g., FLAC, MP3, WAV), upload via a dedicated command that returns a product usable as reference_audio. Specify features to preserve or change (e.g., melody, vocal texture) to guide generation, but results require auditory review.

5. Generation and Task Management: Before submitting a beatra.music.generate request, confirm the production card and obtain single approval. After generation, retain the task_id and poll via beatra.tasks.get, adhering to deadline_at; if not returned, actively poll for 30 minutes then stop. Task list searches must match parameters to avoid duplicate payments.

6. Delivery and Review: On successful tasks, display entries from task.output.clips, providing clip.audio.url, duration, MIME type, etc. Review includes composition, vocal performance, and arrangement, but singer identity, precise duration, etc., are post-generation review attributes, not guaranteed.

Applicable Boundaries and Notes

• Quality Not Guaranteed: Generated attributes like singer identity, pronunciation, and loop points are post-generation review qualities, not promised features.
• Paid Confirmation: Production card approval authorizes only one generation; any parameter modification (e.g., prompts, title, model) requires re-confirmation and a new client_request_id.
• Task Recovery: If a task fails or ID is lost, use beatra.tasks.list to search recent tasks, reusing the ID only if parameters match exactly to avoid automatic retries leading to extra costs.
• Automatic Updates: The client silently checks for updates every 24 hours, with failure not affecting current usage, but not serving as a reason for retries.
• Model Limitations: Only use control items documented for the selected model series; reference audio guides results rather than guarantees them.