Build Skill

When you're writing a new Claude skill and want to avoid the usual three-draft tournament in favor of one deliberate version, this workflow walks you through it. It requires you to establish a failing baseline first, what goes wrong without the skill or what a weak version would miss, then builds a single draft through plan review and targeted refinement. The process is opinionated about writing for senior engineers rather than shell automatons, keeping frontmatter trigger-focused rather than summary-heavy, and pushing back on review feedback that adds ceremony without improving safety. Especially useful when you're motivated by observed agent drift or need to encode decision boundaries that resist rationalization.