The Agent Skills Directory

[COMMAND_EXECUTION]: The script review_prompt.py executes commands defined in the --validator and --smoke CLI flags using subprocess.run(shell=True) in the _run_configured_commands function. This is a core feature for verifying prompt performance but allows for arbitrary command execution based on CLI input.
[CREDENTIALS_UNSAFE]: The variable SCILLM_KEY in review_prompt.py has a hardcoded default value of sk-dev-proxy-123. While likely a dummy value for development, hardcoding credentials remains a security risk.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface because it processes untrusted template and source data, allows an LLM to modify that data, and subsequently executes commands. 1. Ingestion points: Untrusted data enters the context through the --template, --source, and --payload flags in review_prompt.py. 2. Boundary markers: The script uses triple backticks to delimit content in model requests, providing basic isolation. 3. Capability inventory: The skill performs file writes to the template path and executes shell commands via the _run_configured_commands function. 4. Sanitization: The script strips markdown formatting fences from LLM fixes before writing them to disk, but does not further validate the content.

review-prompt