The Agent Skills Directory

[PROMPT_INJECTION]: The skill is designed to process untrusted data such as source code, pull request descriptions, and error logs, which constitutes an indirect prompt injection attack surface.
Ingestion points: The skill explicitly requests the user to provide code, PRs, and error messages in SKILL.md for analysis.
Boundary markers: There are no specific instructions or delimiters provided to the agent to treat the reviewed code as untrusted data or to ignore instructions embedded within it.
Capability inventory: The skill's primary capabilities are limited to generating text-based reviews and reports based on the provided input.
Sanitization: There is no mention of sanitizing or escaping the untrusted input before it is processed by the agent.
[COMMAND_EXECUTION]: The instructions include the use of common development and debugging commands as part of the recommended workflow for the user or agent.
Evidence: SKILL.md and references/nextjs-patterns.md reference the use of git log --oneline for identifying regressions and curl commands for triggering cache revalidation endpoints.
[SAFE]: The skill provides extensive guidance on security best practices, including correct environment variable management, authentication checks, and input sanitization in the reviewed code. All mentions of secrets or credentials use generic placeholders (e.g., REVALIDATE_SECRET, API_SECRET).

code-review-web