An Agent Skill is not passive documentation — it is a set of instructions that an AI coding agent may act on (Claude Code, Codex, Cursor, Copilot, and others) while it has access to your shell, files, and network. A malicious or careless skill could instruct an agent to run destructive commands, exfiltrate secrets, weaken security checks, or take other harmful actions. For that reason, every skill in skillsdeck is reviewed before it is merged, and we ask the community to help us keep the catalog safe.

Skills submitted to skillsdeck must not instruct an agent to:

• Run destructive or irreversible commands without explicit user confirmation (e.g. rm -rf, force-pushes, dropping databases).
• Read, transmit, or log secrets, credentials, tokens, or private user data to any external destination.
• Disable, weaken, or bypass security controls, tests, authentication, or sandboxing.
• Install or execute code from untrusted or unpinned sources.
• Evade detection, obfuscate their behavior, or hide what they are doing from the user.

Skills that violate these rules will be rejected or removed. See the pre-PR checklist in docs/authoring-skills.md before you submit.

Every push and pull request is scanned for hardcoded credentials by gitleaks in CI — this is the authoritative secret gate and blocks the build on any finding (config in .gitleaks.toml). Separately, contributors can run npm run lint:secrets locally: an advisory scan that also flags absolute local paths and private/localhost URLs, which are portability problems a shared skill should not contain. Automated checks assist review; they do not replace it.

If you find a skill in this catalog that behaves unsafely, or a security issue in the tooling (the CLI, scripts, or CI), please report it privately:

• Open a GitHub security advisory on this repository, or
• Email the maintainers (see CODEOWNERS).

Please do not open a public issue for a security report until we have had a chance to review and respond. We aim to acknowledge reports within a few days.

skillsdeck is distributed from the latest commit on the default branch. Fixes are applied to main; there are no separately maintained release branches.