The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The skill is specifically designed to ingest and act upon data from external sources, such as third-party code review feedback, which is an inherent attack surface.
Ingestion points: External review comments processed in references/code-review-reception.md.
Boundary markers: The skill relies on cognitive protocols (skepticism, verification loops) rather than technical delimiters like XML tags or clear-text markers to separate reviewer input from system instructions.
Capability inventory: The skill utilizes git CLI operations and the Task tool for subagent dispatch.
Sanitization: The 'Response Pattern' (READ → UNDERSTAND → VERIFY → EVALUATE) serves as a logical sanitization layer, instructing the agent to evaluate technical correctness before execution.
[Prompt Injection] (SAFE): The skill contains strong, absolute instructional language (e.g., 'The Iron Law', 'Non-negotiable', 'If you lie, you'll be replaced'). These are identified as quality enforcement mechanisms for code-review rigor rather than attempts to override system safety or global constraints.
[Command Execution] (SAFE): The skill encourages the use of standard development tools like git and linter/test execution. These are used for status verification as part of a developer-oriented workflow and do not involve shell injection or obfuscated commands.

code-review