LLM-GENERATED SKILLS WORK, IF YOU GENERATE THEM AFTERWARDS

Summary from the TLDR Dev newsletter

Research shows that LLMs fail at pre-task skill generation by embedding incorrect assumptions. Prompting models to distill procedural knowledge after task completion is better, as this captures iterative insights that can't be captured through initial planning.