The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes standard development tools and package managers for environment setup and model discovery.
Recommends using pnpm, npm, or yarn to install the ai package.
Uses curl and jq to fetch and filter the latest model IDs from a Vercel-managed endpoint.
Employs grep to search for documentation and API definitions within the project's node_modules.
[EXTERNAL_DOWNLOADS]: Accesses official Vercel resources to provide accurate, up-to-date information.
Fetches model metadata from https://ai-gateway.vercel.sh to prevent reliance on hardcoded or outdated model names.
Searches documentation at https://ai-sdk.dev when local files are not present or sufficient.
[PROMPT_INJECTION]: Includes instructions for the agent to ignore its internal knowledge regarding the AI SDK.
This grounding technique is used to prioritize the provided documentation and source code over potentially obsolete training data, which is a defensive measure against halluncinated or deprecated APIs.
[INDIRECT_PROMPT_INJECTION]: The skill has an ingestion surface for potentially untrusted data through documentation lookups.
Ingestion points: Reads local project files (package.json), bundled SDK documentation (node_modules/ai/docs/), and remote search results from ai-sdk.dev.
Boundary markers: No explicit delimiters are specified for the documentation content processed by the agent.
Capability inventory: The skill has access to shell commands (curl, grep), file system reads, and package installation tools.
Sanitization: There are no mentioned processes for sanitizing or validating the content of the fetched documentation before the agent acts upon it.

ai-sdk