The Agent Skills Directory

[DATA_EXFILTRATION]: The skill reads the contents of a user-specified directory, packages them into a compressed tarball, and transmits the data to a remote server. While it excludes standard build artifacts and version control directories, it does not automatically exclude sensitive files such as .env or other local secrets, which could result in their upload to the deployment service.
[EXTERNAL_DOWNLOADS]: The script uses curl to perform a POST request to https://claude-skills-deploy.vercel.com/api/deploy. This endpoint belongs to Vercel, which is a well-known technology provider and service.
[METADATA_POISONING]: The skill's metadata lists 'vercel' as the author, which could be perceived as an official endorsement or origin, despite being provided by a community author. The skill behavior is consistent with the described deployment functionality.
[INDIRECT_PROMPT_INJECTION]: The skill processes project configuration files to determine the deployment environment, creating a surface for indirect influence.
Ingestion points: The scripts/deploy.sh script reads and parses the content of the package.json file found in the project directory.
Boundary markers: None are present; the script directly greps for strings within the file content.
Capability inventory: The skill has the ability to read local files, package them via tar, rename files via mv, and perform network uploads via curl.
Sanitization: The script performs basic substring matching (grep) to identify dependencies, which limits the potential for complex injection attacks but does not sanitize the input for other logic flows.

vercel-deploy