Llm Patterns

This is about building apps where LLMs do the actual business logic, not just generate text. You get patterns for wrapping LLM calls with Zod schemas, organizing prompts as versioned functions, and a three-tier testing approach: mocked unit tests for speed, fixture tests for parsing validation, and eval suites that run nightly in CI to catch accuracy regressions. The core idea is LLM for logic, code for plumbing. It includes cost tracking utilities and shows how to structure projects so prompts live in their own directory instead of scattered through your codebase. Useful when you're building classification systems, extraction pipelines, or anything where the LLM is making actual decisions rather than just filling in templates.