The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied input (the "target" of the review) and interpolates it directly into a prompt template for a subagent call using the Agent tool. This creates a surface for indirect prompt injection where a malicious target description could attempt to subvert the subagent's instructions.
Ingestion points: User-defined "target" and "specialist role" parsed in SKILL.md.
Boundary markers: Absent in the subagent prompt template (e.g., <target> is inserted without delimiters).
Capability inventory: The subagent has access to Read, Write, Glob, and Grep for file analysis, and the Agent tool for further delegation.
Sanitization: The skill mentions input sanitization for generating filenames, but does not sanitize the content injected into the subagent prompt.
[SAFE]: The skill provides instructions for the user to manually override protected file hooks (using CLAUDE_ALLOW_PROTECTED=1) when editing configuration files. This is presented as guidance for the user and does not involve the agent bypassing security controls autonomously.

specialist-review