Claude Skill

It defaults to safe automation with `acceptEdits` mode, skips model flags unless you explicitly want an override, and knows the difference between tool availability and permission approval. The templates cover read-only analysis, edit runs, JSON output, session resumption, and MCP integration. One honest take: the real value is in the permission mode guidance, because `bypassPermissions` sounds tempting but the skill correctly steers you toward explicit allow rules or `dontAsk` unless you're already in a sandbox. It also checks your local `claude --help` output before assuming flag syntax, which matters when the CLI moves faster than blog posts.