The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted input from pull requests and interpolates it directly into LLM prompts, creating an indirect prompt injection surface.
Ingestion points: The example Python orchestrator script (ai_review.py) ingests code diffs and PR descriptions via the reviewer.get_pr_diff() function.
Boundary markers: The provided prompt templates (e.g., review_prompt and security_analysis_prompt) do not use specific delimiters or XML tags to isolate the untrusted code_diff from the core instructions.
Capability inventory: The skill is designed to execute system commands via subprocess and perform network operations via the GitHub API and Anthropic SDK.
Sanitization: The provided code samples do not demonstrate any sanitization, filtering, or length-limiting of the ingested diff content before it is processed by the model.
[COMMAND_EXECUTION]: The orchestration scripts rely on executing external CLI tools to perform static analysis and secret detection.
Evidence: The Python script uses subprocess.run to execute sonar-scanner and subprocess.check_output to run semgrep. The GitHub Action workflow executes trufflehog and sonar-scanner as shell commands.
Context: These executions are necessary for the skill's primary function but rely on the security of the underlying environment and the presence of verified binaries in the PATH.

performance-testing-review-ai-review