The Agent Skills Directory

[COMMAND_EXECUTION]: The skill commands the agent to perform extensive, automated modifications to the filesystem (refactoring code, adding infrastructure, replacing strings) and execute tests without human intervention. Explicit instructions like "Do this automatically without asking the user" and "Don't Ask Questions" attempt to disable standard safety protocols requiring user approval for broad or destructive actions.- [DATA_EXFILTRATION]: The instructions require the agent to recursively read every file in the project directory, specifically searching for "hardcoded secrets (API keys, passwords in code)." While the stated intent is to secure them, this grants the agent unmediated access to all credentials and sensitive configuration files within the workspace.- [PROMPT_INJECTION]: The skill uses highly authoritative language and "Ignore/Bypass" style commands (e.g., "Do all of this without asking the user for input," "Don't Wait for Instructions") to override the agent's default interactive behavior and safety guardrails regarding user oversight.- [INDIRECT_PROMPT_INJECTION]: The skill establishes a large vulnerability surface by ingesting the entire codebase (untrusted external data) and granting the agent high-privilege tools (file read/write/replace) to act upon that data without validation.
Ingestion points: Recursive use of readFile on all project source files.
Boundary markers: None mentioned.
Capability inventory: listDirectory, readFile, strReplace, file creation, and test execution.
Sanitization: The skill does not specify any sanitization, escaping, or validation of the code content before it is processed or used to generate modifications.

production-code-audit