Paul Bakaus argues that skill engineering is more than writing a good prompt. He treats a skill as a reusable workflow that shapes what a coding agent sees, which tools it uses and how it verifies its work.
His examples include an adversarial reviewer that challenges an initial answer, divergent approaches that escape a model's familiar defaults, specialist skills, persistent memory and scripts for deterministic parts of a task.
Bakaus also demonstrates hooks and live browser feedback that bring results back into the agent's loop. He discusses portability across agent harnesses and suggests testing skills against weaker models to expose missing instructions. The talk offers design patterns, not measured proof that every pattern will work in every codebase.
Watch on YouTube
