The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from GitHub pull requests, including titles, bodies, and comments, which could potentially contain malicious instructions intended to influence the agent's behavior. This risk is inherent to tools analyzing external repositories and communications.\n
Ingestion points: scripts/prefetch.sh retrieves PR metadata and comments using the gh CLI, which are then passed to the model via scripts/load-prompt.sh.\n
Boundary markers: The skill uses explicit delimiters and warnings. In scripts/load-prompt.sh, untrusted comment bodies are wrapped in <untrusted_user_comment> tags with instructions to ignore embedded directives.\n
Capability inventory: Worker subagents call bash, python3, node, and gh but operate in read-only or isolated sessions as described in SKILL.md.\n
Sanitization: scripts/load-prompt.sh sanitizes input by stripping control characters and escaping closing tags to prevent block escape.\n\n- [EXTERNAL_DOWNLOADS]: The skill fetches external content and utilizes third-party tools at runtime.\n
Angle Rubrics: The aeo angle fetches additional reference rubrics from the coreyhaines31/marketingskills repository on GitHub at runtime using gh api. This content is used as reference text for the AI audit.\n
Tool Execution: The design angle executes the impeccable CLI tool via npx. While this involves downloading code, it targets a recognized project and the execution is restricted to its audit function. The tool version is unpinned by default ('latest').

woostack-review