The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (pull requests, diffs, and source code) as its primary function. While this presents an attack surface where a malicious payload in the reviewed code could attempt to influence the agent, the skill provides explicit instructions and detection themes (Theme 25) to mitigate this by treating all reviewed content strictly as data rather than instructions.
Ingestion points: SKILL.md and references/pr-context.md define triggers for reviewing external patches and diffs.
Boundary markers: SKILL.md (Step 7) and references/themes.md (Trigger 25.1) provide explicit instructions to treat reviewed content as data and ignore embedded instructions.
Capability inventory: The skill uses bash, view_file, grep, and edit_file to perform audits and apply fixes.
Sanitization: Instructions mandate verification of all findings and use of standardized reporting formats to prevent control flow hijacking.
[COMMAND_EXECUTION]: The skill utilizes powerful tools like bash and edit_file and provides a TypeScript utility (scripts/code-review.ts) to perform local file system audits. These capabilities are directly aligned with its stated purpose of code review and finding remediation, and the skill includes procedures for surgical, evidence-backed changes to minimize risk.

code-review