The Agent Skills Directory

[COMMAND_EXECUTION]: The skill directs the agent to execute shell commands to inspect local system directories.
Evidence: In SKILL.md, the "Companion check" section instructs the agent to run ls on hidden skill directories such as ~/.claude/skills/ and ~/.agent/skills/ to identify installed components.
[REMOTE_CODE_EXECUTION]: The skill facilitates the download and execution of external code for installation purposes.
Evidence: Both README.md and SKILL.md provide instructions for installing companion tools using npx skills add AbsolutelySkilled/AbsolutelySkilled, which downloads and executes scripts from a remote source.
[PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection via project files.
Evidence:
Ingestion points: The agent ingests untrusted code content through git diff and git diff --cached commands during the analysis phase.
Boundary markers: There are no explicit delimiters or instructions provided to the agent to disregard instructions potentially hidden within the source code being reviewed.
Capability inventory: The skill utilizes git, ls, and file read operations across various project files.
Sanitization: No sanitization or filtering of the diff output is performed before the agent processes it for feedback generation.

code-review-mastery