Prompt Engineering with Claude

By the end of this lesson you will know why Anthropic trained Claude to respond well to XML-tagged prompts and the three uses that matter in production code: separating input chunks with <document> / <ticket> tags, shaping Claude's output inside tags like <category> for parse-safe replies, and scaffolding reasoning with <thinking> sections. This is the first technique the Claude AI Engineer path teaches for turning "kinda works" prompts into production-shape ones.

By the end of this lesson you will know when to escalate a Claude prompt from zero-shot instructions to few-shot examples, which specific tickets to pick as your examples (edge cases beat typical cases), and how to structure the example block with XML tags so Claude reliably matches the pattern. Building on the XML-tag work from L1-L2, you'll see how few-shot examples raise Orion Analytics' triage-bot accuracy from ~85% to ~97% on the same 10 tickets.

By the end of this lesson you will know when to add an explicit <thinking>...</thinking> reasoning surface to a Claude prompt (chain-of-thought), how the 5.x-series adaptive-thinking output_config.effort parameter does the same job automatically on Opus 5.5 (and sometimes Fable 5.1), and which one to reach for based on whether you need to log/audit the reasoning or just want higher accuracy on hard multi-step questions. Building on the XML tags from L1 and few-shot from L3, chain-of-thought is the third technique for turning "kinda works" into production-shape.

By the end of this lesson you will know how Claude's prompt caching feature works (cache_control markers + 5-minute TTL), which parts of a prompt are worth caching (system prompt, few-shot examples, reference documents — anything > 1024 tokens that repeats across ≥ 2-3 calls), how to structure a prompt so the cache actually hits, and how to read the cache_read_input_tokens / cache_creation_input_tokens metrics in the response to confirm you got the savings. Combined with the XML tags (L1), few-shot (L3), and CoT (L5) techniques you already know, caching is what turns Orion Analytics' ~200,000-tickets-per-month workload from a $900/month Claude bill into a ~$100/month one.

By the end of this lesson you will know what the Anthropic Prompt Improver is (an in-Playground / API-callable tool at platform.claude.com that takes a prompt + a test set and automatically iterates the prompt against a rubric), when to reach for it (mid-cycle when you've squeezed the manual techniques from L1-L8 but still have a specific accuracy tail), and how to build a good test set (8-20 rows covering typical + edge + adversarial cases). Skipping this step is the single most common reason production Claude prompts leave 3-5 accuracy points on the table.