The Agent Skills Directory

[COMMAND_EXECUTION]: The skill wraps the Vercel CLI to manage cloud infrastructure, using a safety-first approach to command execution.
Evidence: Commands are categorized into four tiers: Safe, Write, Destructive, and Forbidden.
Control: Destructive actions (like production deploys) require explicit user confirmation, and Forbidden actions (like deleting projects) require a triple-confirmation protocol with typed validation. These measures effectively prevent unauthorized or accidental critical changes.
[EXTERNAL_DOWNLOADS]: The skill instructions include installing official software packages from well-known sources.
Evidence: Mentions npm i -g vercel and @vercel/vc-native.
Context: These are official tools from Vercel, a well-known service provider, and follow standard installation patterns.
[PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection by ingesting external data during runtime.
Ingestion points: The skill processes output from vercel logs and fetches documentation from vercel.com/docs to troubleshoot errors (self-healing).
Boundary markers: None explicitly defined for command output processing, which is a common pattern for CLI skills.
Capability inventory: Extensive capabilities through the Vercel CLI as listed in references/commands.md.
Sanitization: The risk of malicious data influencing the agent is mitigated by the skill's mandatory safety tiers, which ensure that no significant action is taken without human oversight.

vercel-cli