Changelog
Every substantive change to the study material, corrections included. When a question here taught something wrong, it gets listed — what it said, and what the documentation actually says.
Certification material goes stale quickly: an answer that was correct last quarter can be a 400 error today. Publishing the corrections is how you can tell whether this site is being maintained or just left up.
Answer bank audited end to end
Every question was re-checked against Anthropic's official documentation. Eleven had wrong or outdated answers and have been corrected, seven were removed as unsupportable, and 40 were added. The bank went from 60 questions to 93 — and every single one now cites the exact documentation line its answer rests on, verified as a literal quote.
Corrections11 questions taught the wrong answer
- A question on the Claude Agent SDK taught "handoffs" as the way to delegate to another agent. That is OpenAI Agents SDK terminology and does not exist in Anthropic's SDK — the answer is subagents. Corrected.
- A Claude Code CI question recommended a --yes flag for permissions. No such flag exists; the correct flags are --permission-mode and --dangerously-skip-permissions. Corrected.
- A permissions question claimed Claude Code has no cross-project scope. ~/.claude/settings.json is exactly that, which left the question with no correct answer. Rewritten around settings precedence.
- Two questions taught assistant prefill and extended thinking with budget_tokens. Both were correct when written and both now return a 400 on current models. Rewritten to test the migration instead.
- A question on MCP server configuration pointed at .claude/settings.json. Project MCP servers live in .mcp.json. Corrected.
- A transport question conflated HTTP+SSE with Streamable HTTP, which superseded it. Corrected.
- A CLAUDE.md question taught that project files take precedence over user files. They are concatenated, and a genuine conflict is resolved arbitrarily. Corrected.
- A context-window question said the API errors when history plus response exceeds the window. On Claude 4.5 and newer the request is accepted and generation stops with model_context_window_exceeded. Corrected.
- A slash-command question taught .claude/commands/. Custom commands have merged into skills; existing files still work but .claude/skills/ is the documented place for new shared ones. Corrected.
- A 529 overload question had no correct answer among its four options — the documented guidance, retry with exponential backoff, was not offered. Rewritten.
- A cost-optimisation question advised deleting few-shot examples. Examples are among the most reliable ways to steer output, and a fixed example block is an ideal prompt-caching target. Re-keyed to caching and model routing.
Other changes
- FixedThe practice exam was solvable without reading it. The correct answer sat at option B in 54 of 60 questions, and was the longest option in 81 of 100 — so "always pick B" scored about 90% and "pick the longest" about 81%. Both are now at chance, and the build fails if either bias returns.
- FixedThe practice exam served the entire question bank on every attempt, so there was no second attempt. It now draws 60 questions from a pool of 100.
- FixedTwelve questions appeared in more than 99% of exams because the scenario selector always chose the same groups. Now randomised.
- FixedThe practice exam ran a 60-minute timer. The real exams are 120 minutes.
- FixedSign-in and registration were broken and had been for some time — both pages failed to load rather than showing a form.
- FixedThe practice exam and diagnostic rendered with dark styling bleeding into light mode.
- Added40 new questions, every one citing the documentation line its answer depends on.
- ChangedSeven questions were removed rather than corrected. Each asserted something no Anthropic documentation supports — a "DAG-based orchestrator" that is not their vocabulary, a single "most important" metric they never rank, a hook rule their own example contradicts. Every question that remains cites the documentation line its answer rests on, and the build now fails if one does not.
- AddedExam results now lead with what you got wrong, grouped by domain, with the explanation shown immediately and a link back into the relevant lesson.
- ChangedExplanations are free. A paywall on them was removed, along with the account requirement to see your results.
Spotted something wrong? Every question cites the documentation line its answer rests on, so a mistake should be checkable. Corrections are welcome.