The Agent Skills Directory

[SAFE]: The skill incorporates a robust safety model designed to audit untrusted repositories securely. It explicitly instructs the agent to treat all repository content as untrusted data rather than instructions, preventing prompt injection attacks from the code being analyzed.
[COMMAND_EXECUTION]: The skill executes local commands such as rg for static analysis and manage.py for runtime checks. These are protected by a mandatory safety gate requiring explicit user approval before execution, mitigating the risk of running code from untrusted repositories.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted project files which may contain malicious instructions. It mitigates this surface by explicitly directing the agent to ignore any in-repo text that attempts to change the audit's scope, scoring, or recommendations.
[DATA_EXFILTRATION]: The skill includes strict data handling rules to prevent the exposure of sensitive information. It mandates the redaction of secrets (API keys, passwords) discovered during the scan and prohibits verbatim output of sensitive code fragments, ensuring project data remains within the user's environment.

django-doctor