Studying for the Claude Certified Architect – Foundations exam? I built an open-source kit: 90 original practice items, and a report that names which kind of wrong answer you keep picking

If you're preparing for CCAR-F — Claude Certified Architect, Foundations — this is an open-source kit for it: 30 task statements, 5 domains, 98 linked wiki notes and 90 original practice items. Rather than list features, here's what its diagnostic actually outputs, so you can judge it before installing anything. A 60-item timed run scoring 77% produces this: distractor family chosen / present rate prompt-instead-of-enforcement 3 / 15 20.0% blames-wrong-component 2 / 20 10.0% suppresses-signal 1 / 13 7.7% solves-different-problem 6 / 82 7.3% unreliable-proxy 1 / 16 6.2% over-engineered 1 / 25 4.0% Every wrong option in the bank is tagged with why it's wrong — one of seven named families. The report ranks them by rate (how often you picked it ÷ how often it was actually on the table), not by raw count. That distinction is the whole point. solves-different-problem has the most raw hits and carries no signal: it's roughly half of all wrong options on any form, so sitting at its base rate means the general skill is intact. Ranking by count would have named it the problem. The real finding is prompt-instead-of-enforcement at ~2.6× its availability — reaching for a prompt instruction where a configuration value already guaranteed the outcome. Here's one of the items that produces that pattern: You have several extraction schemas and the document type is not known in advance. You need to guarantee the model returns structured output rather than a prose reply. Which tool_choice configuration is appropriate? The tempting answer is tool_choice: "auto" plus a system-prompt instruction to always call an extraction tool. The correct one is tool_choice: "any" . Why the tempting one is wrong, verbatim from the bank: it makes a guarantee depend on instruction compliance when a configuration value provides it outright. Every option carries that explanation — the correct ones included. Getting an item right for the wrong reason teaches you nothing, so the explanations are the study material and the score is a byproduct. Second thing that run surfaced: 77% overall, but Domain 2 (Tool Design & MCP) at 55% . A decent average hiding one collapsed domain is exactly what a single percentage can't show you. So the report lists every missed task statement next to the command that fixes it: 2.4 — MCP server integration 0/1 /study 2.4 /quiz --task 2.4 2.3 — tool distribution 1/3 /study 2.3 /quiz --task 2.3 4.5 — batch processing 0/1 /study 4.5 /quiz --task 4.5 And it deliberately refuses to print a scaled score. The exam passes at a scaled 720 out of 1,000 and the raw-to-scaled mapping varies by form, so a fabricated "you scored 743" would invite you to stop studying at exactly the wrong moment. You get percent-correct by domain, which is what the real score report gives you anyway. Disclaimers, up front rather than in a footer: Unofficial. Not written, reviewed, endorsed or sponsored by Anthropic. Contains no exam content. All 90 items are original, written from the published objectives — not even the sample questions printed in the official guide are in there. If you've sat the exam you're under NDA; please don't contribute anything you saw on it. I'm a Claude Ambassador, which is a community program, not an Anthropic role. Saying it because it would be worse to find out later. I'm still studying for this exam myself, so there are no pass-rate claims here. What I can show you is what the tool outputs. It's a git repo you clone rather than a plugin payload, because six of the seven skills write — to your progress, to the question bank, to the wiki — and anything written into a plugin cache is discarded on the next update. Skills: /study , /quiz , /drill , /mock-exam , /progress , /author-question , and /refresh-kb , which re-verifies the wiki against current official docs and logs where the tooling has drifted — because cert material written in July teaches you flags that were renamed in August. MIT for the code, CC BY-SA 4.0 for the content. Repo: github.com/alexiocassanifm/anthropic-certifications The full sample report is at examples/mock-exam-report.md if you want to read the output end to end first. If you're studying a different Anthropic cert, the machinery is shared and certification-aware — adding one means writing a wiki and a question bank, not rebuilding the plumbing. That's the single most useful contribution right now.