The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external data, specifically code diffs and pull request descriptions, which provides an attack surface for indirect prompt injection.
Ingestion points: The ai_review method in the provided Python script processes the diff string, and the Claude review prompt interpolates pr_description and code_diff (SKILL.md).
Boundary markers: The prompt templates use markdown formatting and labels to separate code from instructions, but do not include explicit instructions to the model to disregard instructions contained within the data.
Capability inventory: The skill utilizes the anthropic package for network communication with AI models and subprocess.run for local tool execution (SKILL.md).
Sanitization: No specific sanitization or filtering logic is present for the input data before it is interpolated into the prompt.
[SAFE]: The skill implements a secure approach to external dependencies by warning the agent not to fabricate missing scripts and requiring manual verification of the <support-root>/scripts/ai_review.py file before execution.
[SAFE]: Integration examples for GitHub Actions and Python scripts follow security best practices by referencing environment variables and secrets management for API keys rather than using hardcoded credentials.
[SAFE]: Automated analysis routines use secure subprocess execution patterns with argument lists to mitigate risks of command injection during the execution of static analysis tools like Semgrep and SonarQube.

performance-testing-review-ai-review