The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content originating from git diff outputs and various rule files (e.g., CLAUDE.md).
Ingestion points: Data enters the agent context through the execution of git diff --staged and the reading of project-specific or global rule files (~/.claude/rules/*.md).
Boundary markers: The instructions require reviewers to cite specific rule_source filenames and provide evidence (file:line) for every finding, which helps distinguish between valid code issues and potentially injected content.
Capability inventory: The reviewers are granted Read, Grep, and LSP (Language Server Protocol) access across all scripts. The Edit capability is restricted to Step 4 and is only triggered by explicit user consent.
Sanitization: The skill enforces a "prove-before-report" discipline where reviewers must verify any suspicion found in the diff against the actual files on disk before reporting it.
[COMMAND_EXECUTION]: The skill uses rtk run to execute git diff commands. This is standard behavior for a git-integrated review tool and is limited to gathering information about staged changes within the repository context.

review-staged