Liability adjudication tasks such as legal charge determination require matching event descriptions against domain‑specific skills to decide whether a party bears responsibility. Large language models (LLMs) have shown promise by using executable skills for interpretable reasoning, yet three critical issues remain:\n\n1. Skill libraries become stale as real‑world scenarios evolve;\n2. Skill descriptions often suffer from a semantic‑execution gap, causing LLM reasoning to diverge from expert logic;\n3. Growing skill collections exceed context budgets, leading to omission of crucial skills.\n\nTo tackle these challenges, we introduce KDASO (Knowledge-Data Dual‑Driven Automated Skill Optimization). Its three pillars are:\n\n- Neighborhood‑conditioned induction: automatically generate and refresh skills from recent adjudication cases that are similar to the current query;\n- Execution‑aligned refinement: analyze LLM reasoning trajectories and fine‑tune skill implementations to better align with expert execution;\n- Context‑budgeted routing: dynamically select the most relevant skill subset within a fixed context window, preventing critical skills from being truncated.\n\nAcross ride‑hailing dispute, legal liability, and logical reasoning benchmarks, KDASO consistently outperforms both general‑purpose and domain‑specific baselines. In transfer experiments to concept‑similar scenarios, KDASO achieves an 8.96% performance gain and maintains steady improvements through continuous optimization, demonstrating high transferability and stability of the generated skills.\n\nReview: KDASO’s dual‑driven approach enables adaptive skill evolution and context‑aware selection, offering a practical route for LLMs to deliver interpretable and reliable reasoning in high‑stakes domains such as law.