The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from pull request diffs, metadata, and user comments, creating a surface for indirect prompt injection.
Ingestion points: Untrusted content enters the context via meta.json, diff.txt, and user comments fetched during the prefetch and load stages (scripts/prefetch.sh, scripts/load-prompt.sh).
Boundary markers: The skill uses explicit XML-like delimiters (<untrusted_user_comment>) and includes clear instructions for the AI to treat the block as data-only and ignore embedded directives.
Capability inventory: The skill executes shell scripts, Python, and Node.js code for its logic, and uses the GitHub CLI (gh) to interact with the repository and post comments.
Sanitization: User comments are sanitized by stripping control characters and escaping delimiter tags to prevent prompt escaping (scripts/load-prompt.sh).
[EXTERNAL_DOWNLOADS]: The skill downloads and executes external tools and configuration.
It installs and runs the 'impeccable' design analysis tool from the npm registry via npx (scripts/install.sh).
It fetches audit rubrics from related GitHub repositories (e.g., coreyhaines31/marketingskills) using the GitHub CLI to enhance its review capabilities.
[COMMAND_EXECUTION]: The orchestration logic is implemented through a pipeline of shell, Python, and Node.js scripts. These scripts manage data processing, worker dispatch, and findings finalization. Commands are fixed and operate on internal or sanitized data.

woostack-review