The Agent Skills Directory

[SAFE]: The skill is composed entirely of Markdown files and templates. It contains no executable scripts, binaries, or configuration files that could run code on the host machine.
[PROMPT_INJECTION]: While the skill contains text patterns commonly associated with prompt injection (e.g., 'Ignore all previous instructions'), these are located within defensive guidance sections. Files like rules/memory-and-state.md and rules/safety-and-guardrails.md use these phrases as examples of malicious inputs that developers should sanitize or reject. This is legitimate security documentation and does not pose a threat to the agent's behavior.
[EXTERNAL_DOWNLOADS]: The skill references a wide variety of external URLs for documentation and research (Anthropic, OpenAI, Google, OWASP). These are established technical resources and are referenced neutrally for educational purposes. No automated downloads or remote code execution patterns were found.
[INDIRECT_PROMPT_INJECTION]: The skill addresses this threat category by providing extensive guidance on how to secure applications against it. It recommends using XML delimiters, 'treat as data' instructions, and input classifiers. The skill itself acts as a passive advisor and does not ingest or execute untrusted data at runtime.
[DYNAMIC_EXECUTION]: The skill's observability-and-versioning.md rule suggests invoking other domain-specific skills (e.g., dash0 OTEL skills) to handle specialized instrumentation tasks. This is a standard pattern for skill composition and does not involve unsafe dynamic code generation.

ai-engineering