These examples are deliberately small. Each one has a task contract, a broken starter, a completed solution, and a verifier so you can practice an AI-coding loop without trusting generated code on sight.
- Product slug regression: reproduce a text-boundary bug, ask an agent for the smallest fix, and verify the solution. Standard library only; about three minutes.
- Untrusted payload validation: handle messy camelCase and snake_case inputs, dirty amount strings, and structured error responses.
- MCP tool server: build a standard JSON-RPC 2.0 Model Context Protocol tool server with schema validation and isolated runtime errors.
- Async batch fetcher: control concurrency with Semaphore, back off and retry transient HTTP failures, and collect structured results.
