Think with Skills: Pseudocode Planning for LLM Agents

This method trains reasoning models to generate task-specific pseudocode plans before executing expert-written skills. By explicitly representing dependencies and intermediate results, the approach bridges the gap between declarative skill descriptions and procedural agent actions.