The Agent Skills Directory

[DATA_EXFILTRATION]: Exposure of local environment metadata in documentation files.
Evidence: The file references/composition-pattern.md includes a hardcoded absolute path: /Users/jbm/.agents/skills/vercel-composition-patterns. This reveals a specific username ('jbm') and the local directory structure of the development environment.
[INDIRECT_PROMPT_INJECTION]: The skill establishes a workflow for processing untrusted data from external documentation sources.
Ingestion points: The skill instructs the agent to consult documentation at ui.shadcn.com and base-ui.com, and provides a specific 'conversion rule' in references/overview.md for transforming these external URLs.
Boundary markers: There are no explicit markers or instructions to isolate the agent from potentially malicious content embedded in the documentation it is directed to fetch.
Capability inventory: The agent is permitted to read local skill files, generate UI code samples, and interact with the user based on external information.
Sanitization: Content retrieved from external sources is not validated or sanitized before being integrated into the agent's context.
[METADATA_POISONING]: Inconsistent authorship and installation metadata.
Evidence: The skill manifest identifies the author as microck, while the README.md file suggests an installation path from a different entity, noklip-io/agent-skills. This discrepancy could mislead users about the skill's source and maintenance.

shadcn-base