The Agent Skills Directory

[SAFE]: The skill implements a logical reasoning framework for architectural code review. All file access is limited to the local project workspace and specific plan directories using standard tools like Grep, Glob, and Read.\n- [SAFE]: The 'Parallel Review' and 'Devil's Advocate' processes utilize the platform's native subagent dispatching mechanism (Task tool) to perform complex analysis, which is an expected and safe implementation for this use case.\n- [SAFE]: No network operations, external downloads, or obfuscated code segments were identified in any of the skill's files.\n- [PROMPT_INJECTION]: The skill has an indirect prompt injection surface as it ingests and processes untrusted source code and design plans provided by the user.\n
Ingestion points: Planning documents (e.g., in .claude/plans/) and project source code are read via the Read tool.\n
Boundary markers: The skill does not define explicit delimiters to separate analyzed content from internal instructions, relying instead on the agent's contextual understanding.\n
Capability inventory: The skill is capable of modifying project files (via the Edit tool) and triggering further subagent tasks (via the Task tool) based on the input data.\n
Sanitization: Content from ingested files is interpreted by the agent without prior filtering or sanitization of embedded instructions.

review-design