Agent Skill Builder

Most skills sold online are long prompts saved as a markdown file. Drop one into Claude and nothing happens. No error, no warning, just silence. The reason is almost always the same: no YAML frontmatter, so the runtime never registers a name or a description, and therefore never has any reason to invoke it. There is a second failure that is harder to see and more expensive. The skill loads, runs, and quietly produces worse output than Claude would have produced on its own, while costing you tokens and time. Nobody catches that without an A/B test, and almost nobody runs one, because a skill feels like an upgrade by definition. This skill closes both gaps. You get the frontmatter contract including the invocation controls that keep destructive skills out of the model's hands, the method for writing a description that actually triggers, a decision ladder for when a skill is the wrong answer entirely, and the full eval method: run the same task with the skill and without it, grade both, compare them blind, and be willing to conclude the skill should be deleted. Then it makes the skill cheap to run. Context forking, file handoff between steps, and parse-time substitution took one real production workflow from 51k tokens per run to under 8k with no change in output. It also runs in audit mode. Hand it a skill that never fires and it works a fixed diagnostic order, most severe first, and tells you which problems it fixed and which need your decision.