When many tools and many Skills coexist, source-of-truth conflicts, trigger ranking, and frontmatter dialects become unavoidable: why sharing one folder is not the same as sharing one contract.
A practical companion to Best Practices: how to narrow descriptions, layer SKILL.md vs references, and use checklists plus a versioning strategy to keep quality stable over time.
A real refactor walkthrough: freezing pagination mapping, date-range checks, query fold behavior, org permissions, and export handling as a project Skill so the model’s first pass needs fewer rework rounds.
This post only answers “what deserves a Skill”: the repeat / stable / executable bar; a scenario checklist with counterexamples; a crosswalk to the DeepLearning.AI short course and Anthropic’s official notes. Series context is in parts 1–3.
Use skill-base / skb to publish, search, install, and update team Skills like npm—so you stop copy-pasting rules and fighting version drift.
How Skills load in layers at runtime, what a skill directory looks like, and a copy-paste minimal `SKILL.md`; team distribution and versioning are in part 3.
From “I’ve heard of it / installed a few but can’t explain it” to writing clearly and splitting concerns: what a Skill is, what problems it fits, how to pick topics and author them, plus fragmentation and team rollout.