The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes a local utility at .claude/scripts/code_graph to analyze codebase structure and trace dependencies.
[PROMPT_INJECTION]: The skill is subject to indirect prompt injection risks due to its data ingestion surfaces. 1. Ingestion points: The skill reads untrusted data from the $ARGUMENTS variable and from project files that may contain injected content. 2. Boundary markers: The <tasks> tag is used to delimit task-specific arguments. 3. Capability inventory: The skill can execute local scripts, perform file system operations, and trigger internal agent slash commands like /plan and /code. 4. Sanitization: While it lacks traditional input filtering, it defines strict Red Flag Stop Conditions that halt execution and require user intervention when security-sensitive code (e.g., auth, crypto, PII) is encountered.
[SAFE]: The skill implements rigorous operational guardrails, including confidence-based stopping (below 60%), blast-radius limits (20 files), and mandatory evidence-based pattern matching before modifications are allowed.

cook-auto