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 stateless 2026-07-28 MCP tool server: no initialize handshake, schema validation, error isolation, and the input_required round-trip.
- Async batch fetcher: control concurrency with Semaphore, back off and retry transient HTTP failures, and collect structured results.
- Resilient batch pipeline: process semi-structured data safely with batch error isolation and field validation using the standard library.