The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses authoritative language ("CRITICAL — Your training data is outdated", "do not guess") to instruct the agent to disregard its internal training data in favor of provided content. This content includes fictional model versions (e.g., GPT-5.4, GPT-5.6) and future dates (March 2026) presented as authoritative facts.
[INDIRECT_PROMPT_INJECTION]: The skill documents patterns for building AI-powered features that ingest untrusted user prompts.
Ingestion points: The prompt parameter in generateText, streamText, and generateImage examples in SKILL.md.
Boundary markers: None present; the code examples interpolate user messages directly into the model call.
Capability inventory: Network operations via the Vercel AI Gateway API and associated provider endpoints.
Sanitization: The examples do not demonstrate validation, escaping, or filtering of the input prompt data.
[EXTERNAL_DOWNLOADS]: The skill references and fetches resources from Vercel's official documentation and GitHub repositories.
[COMMAND_EXECUTION]: The instructions include shell commands for environment setup and dependency management, such as vercel env pull and npm install @ai-sdk/gateway.

ai-gateway