The Agent Skills Directory

[PROMPT_INJECTION]: The skill employs 'CRITICAL' and 'IMPORTANT' advisories to instruct the agent to disregard its internal training data in favor of external documentation and the skill's own content. It presents 'future-dated' information (April 2026) and references non-existent models like 'GPT-5.4' and 'Claude 4.6', which acts as a form of knowledge poisoning.
[COMMAND_EXECUTION]: Provides instructions for executing shell commands involving environment configuration (vercel env pull), package installation (npm install @ai-sdk/gateway), and API interactions using curl. These are standard for the described service but require user execution of CLI tools.
[EXTERNAL_DOWNLOADS]: References external resources and package installations from Vercel-controlled domains and official registries like NPM. These are well-known services, but the skill directs the agent to treat these external sources as authoritative overrides.
[CREDENTIALS_UNSAFE]: Discusses the management of sensitive tokens such as VERCEL_OIDC_TOKEN, AI_GATEWAY_API_KEY, and ANTHROPIC_AUTH_TOKEN. While it suggests standard security practices like using .env.local and OIDC, it involves the handling and exposure of authentication secrets in the shell and environment files.

ai-gateway