The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data in the form of code modified during the development session.
Ingestion points: The skill reads modified source code using git diff and git status.
Boundary markers: The instructions do not explicitly include boundary markers for the ingested code diffs.
Capability inventory: The skill can read local files via git, invoke sub-agents via the Skill tool, and propose file edits.
Sanitization: No explicit sanitization or instruction-ignoring markers are defined for the code being reviewed.
Risk mitigation: The skill mandates a "Propose, don't apply" workflow, requiring explicit user approval before any code changes are executed, which effectively mitigates the risk of malicious instructions in the code causing unauthorized changes.
[SAFE]: The skill references 'vercel-react-best-practices' for performance reviews. This reference to a well-known service's guidelines is informative and does not introduce security risks.
[SAFE]: The skill uses standard development tools like git to scope its operations to modified files, which is a best practice for limiting the impact of automated refactoring tools.

do-cleanup-web