Download the study guide as a polished PDF booklet or EPUB.
What's new in v1.1.0
This release corrects two facts that could cost you exam points, adds three new sections and a practice-exam appendix, and freshens smaller details throughout. If you studied from v1.0.x, the corrections below are worth a re-read even if you skip the new material.
Corrections worth re-reading
- MCP server scope precedence was reversed. When the same server name exists at multiple scopes in Claude Code, the order is local > project > user — your local definition overrides the team-shared
.mcp.json, not the other way around. (Section 7) - Assistant prefill is now a legacy technique. Current Claude models reject a request that ends in a partial assistant message. The guide now teaches the replacements — structured outputs, schema-backed tools, and system-prompt instructions — with a migration table. (Sections 1 and 6)
- Smaller fixes: a mismatched tool name in the Section 1 extraction example, and context-window references rephrased to stay accurate as models evolve.
New sections
- Model Selection and Inference Controls (Section 12) — choosing between Haiku/Sonnet/Opus tiers per pipeline step, cheap-to-expensive routing, adaptive thinking and the effort setting, streaming, the full stop-reason table, API rate limits, and token counting.
- Prompt Caching (Section 13) — the exact-prefix-match rule, ordering prompts for cache stability, the silent invalidators that quietly zero out your hit rate, breakpoint placement and economics, and when caching is the wrong lever.
- Security and Trust Boundaries (Section 15) — prompt injection and its architectural defenses, the private-data / untrusted-content / outbound-channel risk triad, vetting MCP servers and hooks like dependencies, and keeping secrets out of context.
Batch Processing is now Section 14 and the Quick Reference Cheat Sheet is Section 16.
Practice scenarios
A new appendix with 15 original multiple-choice scenarios — one per section of the guide, written in the exam's style — followed by an answer key whose rationales explain why each rejected option fails. Use it as a diagnostic to find the sections worth re-reading.
Expanded throughout
- Where tools run — client-side vs server-side vs MCP — and why placement sets the trust boundary (Section 2).
- API-native context management: compaction and context editing vs application-level strategies (Section 5).
- Skills in Claude Code: progressive loading and how they differ from slash commands, rules, and CLAUDE.md (Section 10).
- Cheat sheet, study order, and practice drills updated to match, plus 13 new official documentation links in the resources list.