Published
Security checklist for new skills
Security checklist for new skills
Several teams asked the same question after last week’s chaining post: what do we verify before a skill touches production data?
The five-minute pass
- Read
allowed-tools(or equivalent) inSKILL.md— does it match what you expect? - Run it on a throwaway branch or sample repo first.
- Confirm scripts don’t echo secrets into the chat log.
- Prefer skills that describe outcomes, not open-ended “do whatever” instructions.
- Keep humans on the hook for deploy and payment actions.
For the full framing, see Agent skills best practices.
Catalog note
Skills under the security category were reviewed this week for clearer guardrails in their descriptions. Filter by tag in the catalog if you’re hardening a workflow.