“Coding chat” and “agents that act consistently in your environment” differ by one layer: structured tool docs. Many stacks use a skills folder + SKILL.md (or similar) describing when/how to invoke commands; the runtime injects that into the system prompt.
Four things readers should know
- Skills aren’t magic—they’re instructions plus guardrails (binaries, env vars, config).
- Treat third‑party skills as untrusted—they enter the prompt; malicious text can steer behavior.
- Official repo vs community registry differ—bundles change with releases; community listings need maintainer checks.
- In hybrid local/cloud, inject secrets per your security policy—don’t paste keys into chat for convenience.
If you already run an agent stack, spend thirty minutes on the official “skills/tools” docs—it prevents more incidents than chasing every new model drop.