What problem does it solve? AI-generated vocals sound machine-made because they are perfectly in tune, dynamically flat, and identical across chorus repeats, and prompts like "sing louder" do nothing in generation models. This Skill provides the vocabulary and rules for directing how a vocal should be delivered, not just what notes it sings. ## Core Features & Use Cases - Delivery Vocabulary: Replace ineffective instructions with prompt-ready terms like "brighter", "breathy", "harder consonants", and "smiling delivery" that Suno, YuE2, and MiniMax actually respond to. - Anti-Perfection Pitch Policy: Define where pitch deviation is deliberately preserved (based on arrangement density, not section names) so generated vocals avoid the "AI flavor" of total tuning accuracy. - Four Harmony Types & Doubles: Choose among full-voice upper harmony, airy close harmony, shadow-like lower harmony, and call-and-response phrases, each with its own delivery and EQ strategy, plus single/double/triple panning rules. - Use Case: When a generated chorus vocal sounds buried, first check pitch (flat vocals never cut through regardless of fader level), then write a per-section delivery and chorus-variation plan into the ARR-SPEC vocal block to satisfy lint checks #15 and #16. ## Quick Start Ask the agent to write the vocal block of an ARR-SPEC for your song, specifying the singer persona, per-section delivery, dynamics curve, chorus variations, pitch policy, and harmony stacking plan.
