skill-builderlisted

# Skill Builder — the principles behind a skill that triggers and stays focused > For interactive scaffolding, evaluation and benchmarking of a skill, use Anthropic's official **Skill Creator** (claude.com/plugins/skill-creator). This skill is the complementary **judgment layer**: the opinionated principles that make a skill trigger reliably and stay focused, plus a review checklist and the packaging path. Use it to design, review, or debug a skill, not to generate boilerplate. ## The root virtue: predictability A skill exists to pull determinism out of a stochastic system. What you make repeatable is the **process**, not the output — the agent should take the same approach every run, even when the result differs. Design the procedure; don't script the answer. Two tests fall out of this: - **Completion criteria the agent can check.** The skill must give it a way to tell *done* from *not-done* — otherwise it stops early and calls it finished. - **One leading word.** Anchor the skill on a single concept the model already knows well (e.g. *tracer bullet*, *receipt*, *lesson*, *seam*) and repeat it in the name, the description and the body. A word the model thinks with anchors both *triggering* and *execution* — far more than a phrase it has to decode. ## First: is a skill even the right tool? Claude Code has five extension primitives. Pick the one that fits before writing anything. - **Skill** — a reusable workflow or body of knowledge, pulled in **when its description ma