Dreaming the Right Skill: Capability-Guided Retrieval for LLM Agents

Retrieving reusable skills can improve LLM agents by providing procedural guidance for complex tasks. Existing retrieve-and-rerank pipelines match an agent's query directly to skill documents. As skill libraries grow, query-based matching can return many semantically related skills that provide no useful guidance for the current operation. We introduce DreamSkill, which uses a Dream—a short, query-conditioned description of the required capability, including its operations, intended outcomes, and explicit constraints—to guide both supervision construction and online retrieval. By decomposing compound requests into capability queries and matching each Dream to a supporting skill, we construct 260,631 step-level training pairs. A shared encoder learns to match Query+Dream to full skill documents and uses the same input format at deployment. During execution, the agent independently assesses each candidate against the task, observed state, and available tools before selecting among applicable procedures. One trained retriever serves different executors, each using its own base model for Dream generation and contextual selection. Across three benchmarks and three executors, DreamSkill achieves the highest mean task accuracy or success rate in all nine settings. Its retriever also leads all six usefulness metrics on skill pools of 6,660 and 35,059 documents. On Terminal-Bench, it improves Qwen3.5-9B accuracy by 5.99 percentage points over the strongest skill baselines. On Spider2-DBT, it raises DeepSeek-V4-Flash success from 44.79% without skills to 56.25%.