AI music generators filter artist and song names, and name-dropping prompts poorly anyway. This is the layer that de-references a named song into its musical DNA (key, tempo, structure, instrumentation, production, vocal character) and blends two or more into one prompt.
Fully specified and ready to build.
The decisions I locked in
- A hybrid blend engine. Deterministic rules enforce coherence: they resolve key, tempo and structure instead of averaging two songs into sludge. One LLM call does the creative fusion and writes the final prose.
- MCP first. The first user is an AI assistant calling two tools,
get_song_dnaandblend. Selling to the generator companies comes later. - Start with knowledge, not signal processing. v1 is about 50 famous, well-documented songs with descriptors written by an LLM and refined by hand. Audio analysis waits until the canon needs to scale.
MVP scope
In: the curated canon, the two tools, the hybrid engine, a Suno prompt formatter, and a live demo where the blended prompt produces a track that actually sounds like the blend.
Out, on purpose: analyzing arbitrary tracks, audio feature extraction, an automated validation loop, and formats for other generators.
Pairs with songprint: one handles the sound, the other the words.