The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides detailed rules and examples for code refactoring and performance optimization. This creates an attack surface where an agent, when applying these rules to an untrusted codebase, could be influenced by malicious instructions embedded in the target code's comments or string literals (e.g., instructions attempting to subvert the refactoring logic).
Ingestion points: User-provided React and Next.js source code files during refactoring tasks.
Boundary markers: None detected in the instruction text to differentiate between the skill's rules and the target code's content.
Capability inventory: File system access and potentially shell execution are typical for agents performing the described refactoring tasks.
Sanitization: No explicit sanitization or validation of the input codebase is described in the guidelines.
[METADATA_POISONING]: There is a discrepancy in the author metadata. The SKILL.md frontmatter and metadata.json claim the author/organization is 'Vercel' or 'Vercel Engineering', while the skill is provided in a context associated with 'makerjackie'. Although the README.md transparently attributes the original work to Vercel engineers, the inconsistency in the platform metadata fields could be misleading regarding the official status or support of the skill.
[EXTERNAL_DOWNLOADS]: The skill's instructions and code examples reference several external Node.js packages and official documentation resources. These include well-known packages like swr, lru-cache, lucide-react, and @mui/material, as well as vendor-specific tools like @vercel/analytics. References are primarily to trusted domains including react.dev, nextjs.org, and vercel.com.

vercel-react-best-practices