Skill best practices

Write skills that are clear, least-privilege, and safe under OpenBot approvals.

• State when to use the skill and when not to.
• Mark untrusted data (webhook excerpts, attachments) explicitly in prompts.
• Prefer connector-backed auth over embedding tokens in SKILL.md.
• Keep scripts idempotent; assume approvals may deny a tool mid-run.
• Test deny paths: a skill should fail closed, not invent credentials.