What problem does it solve? Turning a structured music arrangement specification (ARR-SPEC) into working inputs for AI music generation backends is error-prone: each backend (Suno, YuE2, MiniMax, ACE-Step, Stable Audio, symbolic MIDI) accepts different fields, honors different levels of control, and silently ignores what it cannot execute. This Skill provides a single compiler that converts one ARR-SPEC into per-backend inputs without losing information or mis-scoring backends for fields they never claimed to support. ## Core Features & Use Cases - Backend capability matrix (honors contract): Each backend declares exact/hint/none support per ARR-SPEC field in backends.yaml, so unsupported fields are excluded from prompts and from scoring rather than penalized. - Style prompt compilation: Converts structured fields (mood, tempo, key, instrumentation, groove, vocal persona) into prose prompts with hard rules against empty adjectives and raw numeric values. - Section-tag and lyric mapping: Maps form sections to backend tags (e.g., 大サビ to [Bridge]), enforces tag count equals form section count, and chunks lyrics to match. - Degradation and seed strategy: Records per-backend field degradation in render.<backend>, and defines fixed-seed A/B versus best-of-N workflows. - Use Case: You have a finished City Pop ARR-SPEC and want to generate it on both YuE2 (exact ABC control) and Suno (prose prompt). The compiler produces the ABC score path for YuE2, a prose style prompt with section tags for Suno, and logs which fields were degraded for Suno. ## Quick Start Compile my finished ARR-SPEC into Suno and YuE2 inputs, applying the honors contract and writing the degradation records back into the spec.