Skip to main content

Build prompts like product code

prompt-fn turns LLM prompts into type-safe async functions. Every prompt has:

  • Schema-first inputs/outputs powered by Zod (inferred TypeScript types at compile time, validated at runtime).
  • Flexible rendering via template literals or Eta templates for advanced logic/partials.
  • Provider-agnostic models by wrapping Vercel's AI SDK (generateObject/generateText) so OpenAI, Gemini, Anthropic, and Ollama work the same way.
  • Resilience tooling such as pino logging, structured output recovery, and fallback parsing for providers that return malformed JSON.
  • Read the Getting Started guide to install the package with Bun or npm and wire up your first prompt.
  • Explore the Guides to learn how templating, logging, and recovery hooks fit together.
  • Dive into Examples that mirror the real examples/ directory in the repo.
  • Review the API Reference generated with TypeDoc and aligned with the DefinePromptConfig types.

Project goals​

  1. Prompts as functions: const ticket = await definePrompt(...)(input)—no hand-managed strings.
  2. Type safety everywhere: input/output schemas live beside the template.
  3. Engine agnostic: swap providers by passing any Vercel AI SDK model instance.
  4. Operational visibility: pino loggers and recovery utilities keep production runs observable.