What problem does it solve? Researchers often have a clear problem but only a fuzzy technical route, leading to bloated proposals, contribution sprawl, or drift away from the original bottleneck. This Skill turns a vague research direction into a problem-anchored, implementation-ready method plan through structured multi-round external review. ## Core Features & Use Cases - Problem Anchor Freezing: Extracts an immutable bottom-line problem, bottleneck, non-goals, and success criteria that every revision round must preserve, with explicit drift detection. - Iterative GPT-5.5 Review Loop: Sends proposals to a Codex-hosted reviewer that scores 7 dimensions (problem fidelity, method specificity, contribution quality, frontier leverage, feasibility, validation focus, venue readiness) and repeats until score >= 9 or 5 rounds. - Checkpoint Recovery: Persists state to REFINE_STATE.json after each phase so interrupted sessions resume from the last completed phase with reviewer thread continuity. - Use Case: A researcher says "refine my approach for long-context video understanding" and receives a full proposal with system graph, training recipe, minimal claim-driven validation, score history, and a final report ready to hand off to experiment planning. ## Quick Start Ask the assistant to refine your research plan by describing your problem and rough approach, for example: refine my approach to reducing hallucination in retrieval-augmented generation.
