The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute the claude CLI via shell commands to process review prompts.
[INDIRECT_PROMPT_INJECTION]: The review process aggregates untrusted repository artifacts into a prompt, creating a surface for injection.
Ingestion points: The skill reads repository files, excerpts, and diffs to construct a 'brief' for analysis as described in SKILL.md.
Boundary markers: Explicit instructions require the inclusion of a read-only boundary marker: "Read-only review. Do not edit files or run state-changing commands."
Capability inventory: The skill uses shell and CLI tool calling but includes hardening restrictions like --permission-mode plan and explicitly disallows file-writing tools (Edit, Write, NotebookEdit).
Sanitization: The skill mandates a verification phase where the agent must check all model claims against the actual source code and tests before reporting results to the user.
[DYNAMIC_EXECUTION]: To prevent execution hangs during complex reviews, the skill recommends managing the review process via a Python subprocess runner with explicit timeouts and process group termination logic.

fable-review