The Agent Skills Directory

[COMMAND_EXECUTION] (HIGH): The skill provides the agent with specific command recipes for high-impact infrastructure operations. These commands allow for direct modification of production environments with no human-in-the-loop validation.
Evidence: Templates for gh pr merge --squash, vercel env add KEY production, and stripe -p production products create.
[DATA_EXFILTRATION] (MEDIUM): The skill includes commands designed to extract sensitive configuration data and write to local environment files, exposing them to the agent context.
Evidence: vercel env ls | grep production lists production secrets, and echo '...' >> .env.local modifies sensitive local configuration files.
[PROMPT_INJECTION] (HIGH): The skill establishes an indirect prompt injection surface by bridging untrusted data (like PR descriptions or env values) to powerful CLI tools.
Ingestion points: The value in environment variable templates and title/body in GitHub CLI commands.
Boundary markers: Absent. No delimiters or instructions to ignore embedded commands are provided.
Capability inventory: Subprocess execution of gh, vercel, stripe, sentry-cli, npx, and curl.
Sanitization: Absent. The templates use printf or quotes which do not prevent high-level instruction injection in the processed text.
[EXTERNAL_DOWNLOADS] (LOW): The skill relies on npx for executing certain tools, which involves runtime package acquisition from external registries.
Evidence: npx convex env set downloads and executes the convex package at runtime.

cli-reference