The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructs a sub-agent to adopt a specific persona ("senior staff engineer in a bad mood") to ensure a rigorous and adversarial code review. This use of role-play is functional for the skill's primary purpose and does not attempt to bypass core safety guardrails or extract system prompts.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to execute standard git commands (git diff, git log, git status) and utilizes Read and Grep for file inspection. These operations are limited to gathering the context necessary for reviewing code changes in a repository.
[SAFE]: The skill possesses a surface for indirect prompt injection because it processes untrusted data (code diffs and commit messages). The implementation includes mitigations such as instructing the reviewer to be suspicious and explicitly forbidding the sub-agent from editing code or writing files.
Ingestion points: The outputs of git diff, git log, and git status are passed directly into the sub-agent prompt.
Boundary markers: The reviewer prompt uses clear labels like Diff:, Commits:, and Untracked: to separate instructions from the data being analyzed.
Capability inventory: The skill has access to Bash, Read, Grep, and Agent tools. The sub-agent prompt explicitly restricts write capabilities ("Modify no code, write no files").
Sanitization: There is no evidence of specific escaping or sanitization of the git output content before it is processed.

critique