01·Claude quirks: XML, prefill, system field

Introduction

A lesson on Anthropic Claude specifics: why XML tags work better than markdown headers, the <thinking> pattern vs OpenAI reasoning tokens, the assistant-turn prefill as the strongest output-control lever, separate system field, no JSON mode (tool use instead), citations API, computer use, multi-turn conversation in a stateless API, per-model rate limits, deployment via Bedrock and Vertex AI, beta headers, model versioning with aliases, anti-patterns for devs with an OpenAI background, when NOT to choose Claude.