Key takeaways

The through-line

Building well with Claude is specification engineering, not incantation. Every technique in this course — prompts, schemas, tools, caching, agents — is a way of stating what you want precisely enough that the model's capability lands on your task, and a way of verifying that it did.

The ideas that carry

• Choose models by evidence, from the top down. Prototype on the most capable model to learn the ceiling, freeze a small evaluation set, then downshift route by route where quality holds. A production system uses several models, not one.
• The API is stateless, and that is the feature. You control every byte of context on every request — which is what makes history management, injection of fresh data, and caching possible at all.
• The system prompt is a job description, kept frozen. Role, grounding rules, output contract, boundaries — structured with tags, calibrated rather than shouted, and byte-identical across requests so the cache can do its work.
• Reasoning is a dial, not a trick. Adaptive thinking plus the effort parameter replaced "think step by step" — and thinking spends from max_tokens, so budget for it.
• Structured output is guaranteed, and guaranteed output is testable. Schemas via output_config.format and strict tools ended the parse-and-pray era.
• Tool descriptions steer; harnesses gate. Claude decides tool calls from descriptions that say when; your code decides what actually executes — least privilege, confirmation on anything irreversible, errors returned as data.
• Caching is an exact prefix match. Stable content first, volatile last, marker at the boundary — verified in usage, never assumed. It is the largest cost lever you have.
• Check stop_reason before consuming content. Truncation and refusal are normal states that need designed paths.
• Agents are a decision, then guardrails. Four questions (complexity, value, viability, cost of error) decide whether to build one; caps, logging, and evidence-backed claims decide whether it can be trusted.
• MCP turns M×N integrations into M+N — and every connected server is a trust grant to scope deliberately, because tool results are untrusted input.
• Everything a production failure teaches goes into the eval set. Twenty real cases in CI convert "we think it's fine" into "we checked."

What you can now do

Pick and re-pick models with an evaluation set instead of a feeling. Write system prompts that steer current models. Get schema-guaranteed output into typed code. Hand Claude tools safely and run the agentic loop — by hand or with the runner. Cut input costs with cache discipline and batch pricing. Ground answers in your documents with citations. Drive Claude Code with CLAUDE.md, plan-first workflows, and feedback loops. Ship behind a checklist instead of a hope.

If one thing sticks

Measurement beats vibes. The teams that win with Claude are not the ones with secret prompts — they are the ones who can tell, in minutes and with evidence, whether any change made things better or worse.