The Agent Skills Directory

[COMMAND_EXECUTION]: The helper script "scripts/claude_review.py" executes the external "claude" CLI tool using "subprocess.run". Although it passes arguments as a list to mitigate shell injection, the skill's core functionality depends on executing external system binaries.
[PROMPT_INJECTION]: In "SKILL.md", the instructions guide the agent to "request escalation for the helper command" (sudo) if authentication fails in restricted or sandboxed environments. This is a directive for privilege escalation that encourages the agent to seek higher system permissions.
[PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection. The "build_prompt" function in "scripts/claude_review.py" interpolates "review_input"—which can contain untrusted file content or user-provided text—directly into the final prompt sent to the LLM without using delimiters, boundary markers, or explicit isolation instructions.
Ingestion points: Untrusted data enters via stdin in "scripts/claude_review.py", which is filled with project files, diffs, or plans provided by the user or project environment.
Boundary markers: Absent; the content is placed directly under a text label without separators.
Capability inventory: The script's output (verdict, issues, loop_signal) determines the agent's logic flow, including whether to modify artifacts or continue the review loop.
Sanitization: No escaping or validation is performed on the input content before it is embedded in the prompt.

claude-review