The Agent Skills Directory

[SAFE]: The skill is a meta-documentation project intended to guide developers in creating secure and reliable AI agent harnesses. It emphasizes security axioms such as "security through containment," "least privilege," and "fail-closed" defaults.
[EXTERNAL_DOWNLOADS]: The documentation provides links to research papers (arXiv), security articles (Snyk), and official repositories from trusted organizations including Anthropic and Vercel. These references are used neutrally to support the design principles.
[COMMAND_EXECUTION]: The skill contains various shell command examples (e.g., git, claude CLI, gh, pnpm) within code blocks. These are provided as instructional snippets for configuring CI/CD pipelines and developer workflows, not as malicious payloads.
[INDIRECT_PROMPT_INJECTION]: The skill explicitly designs for the mitigation of indirect prompt injection by recommending that agents treat external data (such as pull request titles or git diffs) as untrusted and requiring structured validation or human-in-the-loop gates before processing.
[DYNAMIC_CONTEXT_INJECTION]: In references/skill-authoring.md, the skill documents the !command syntax used by some agent platforms to inject shell output into the skill at load time. This mention is purely descriptive, explaining platform features to the user, and is not an attempt to execute malicious commands.

agent-native-project-design