What problem does it solve? Developing theorem candidates in machine-learning theory requires turning an approved research perspective into one concrete, non-duplicate (setting, goal) idea that downstream formalization can act on without inventing the core research choice. This Skill performs that branch-local idea-generation step while avoiding duplicates and incorporating failure diagnostics from prior proof reviews. ## Core Features & Use Cases - Branch-local idea generation: Produces one concrete (setting, goal) candidate under a fixed parent perspective and writes it to perspective_M/idea_N/idea.md using a shared template. - Duplicate and failure avoidance: Reads theory_tracker.md and prior same-perspective idea and review artifacts to reject near-duplicates and candidates that repeat known theorem-level obstructions. - Failure-driven repair: Consumes proof_sketch_review.md, global_proof_review.md, and aggregated proof_review.md diagnostics to apply minimal, target-preserving repairs, or returns NO_VIABLE_IDEA when no candidate survives the hard gates. - Use Case: After a proof sketch review returns IDEA_FAIL for a generalization-bound branch, rerun this Skill to generate a revised idea that repairs only the implicated assumption while preserving unaffected structure. ## Quick Start Ask the agent to run the subagent-idea-generator for perspective 2, idea 3, using the assigned perspective from Perspective_Selection.md and the triggering proof_sketch_review.md.
