Key takeaways

The through-line

Building well with Claude is specification engineering, not incantation — and it is bigger than the API. The course's arc: speak the platform's native shapes precisely, route work to the right surface, wrap capability in harnesses that verify and contain it, and then re-draw the development lifecycle with agents at every stage and humans at every judgment gate.

The ideas that carry

• Choose models by evidence, from the top down — prototype at the ceiling, freeze an eval set, downshift route by route. A production system uses several models, not one.
• Pick the surface, not just the model. Apps and Projects for thinking, Claude Code for the codebase, Chrome for the browser tail, API for your product, Agent SDK for your own agents — fluency transfers across all of them.
• The system prompt is a job description, kept frozen and versioned like code. Reasoning is a dial (effort), not an incantation.
• Structured output is guaranteed, and guaranteed output is testable — and the schema itself steers: names, enums, and field order are prompt design.
• Tool descriptions steer; harnesses gate. Errors are data; parallel results return in one message; strict schemas by default.
• Caching is an exact prefix match — stable first, volatile last, verified in usage. The largest cost lever you have.
• Check stop_reason before consuming content. Truncation and refusal are normal states with designed paths.
• Evidence goes in front of the model — images and PDFs as blocks, citations on from day one, long-context-plus-caching before RAG, tools-as-retrieval in between.
• Computer use is API-last, pixels-as-fallback — sandboxed, briefed precisely, gated on anything irreversible.
• Context is a budget: edit the stale, compact the old, persist the important, isolate the parallel — the doctrine behind subagents and multi-agent teams alike.
• Exhaust the five workflow patterns before building an agent; four questions decide when an agent is earned.
• MCP turns M×N integrations into M+N, and every server is a trust grant.
• Claude Code compounds when you specialise it — CLAUDE.md, skills, hooks, subagents, worktrees — and headless mode plus the Agent SDK make the same harness ambient in CI and embeddable in products.
• The ADLC is the synthesis: agents at every lifecycle stage, artifacts over actions, autonomy earned per-stage by eval evidence, humans at the named judgment gates.
• Injection is defended in layers, never solved — data marked as data, least privilege, gates, audit; sandboxes contain what executes.

What you can now do

Route any piece of work to the right Claude surface. Ship API features with schemas, tools, caching, and streaming done properly. Ground answers in documents, images, and live screens. Build workflows, single agents, and multi-agent teams with contexts that stay sharp. Turn Claude Code into your team's specialised instrument and run it headless across the lifecycle. Hold all of it to an eval suite and a defence-in-depth security posture — and know exactly which canonical source to check when the platform moves.

If one thing sticks

The teams that win with Claude are not the ones with secret prompts. They specify precisely, verify relentlessly, and build harnesses that make capability safe to use at speed — so every model improvement lands as pure upside.