| name | review-bank |
| description | Adversarially review an Anthropic University question bank for wrong keys, ambiguity, tells, weak distractors, and duplication (contributor tool). Use before opening a PR that changes questions. |
| argument-hint | <exam-id> [domain-id] |
| allowed-tools | Read, Grep, Glob, Edit |
Review a bank (contributor)
Audit content/<exam-id>/questions.json (optionally one domain) before it ships.
Verify keys against the docs
Confirm every keyed answer with the official docs (claude-code-docs MCP server / claude-api
skill). Re-check details exams love to test: stop_reason values, batch API (~50% / up to 24h),
-p/--print, --output-format json, .claude/rules/ glob scoping, skill frontmatter
(context: fork, allowed-tools, argument-hint), .mcp.json scopes and ${ENV} expansion,
prompt caching, plugin/marketplace layout.
Checklist (most severe first)
| Category | Look for |
|---|
| wrong-key | Keyed answer is not best/only correct; another option also correct |
| ambiguous | Two options defensibly correct; stem under-specified |
| tell | Correct answer obvious from length, specificity, or grammar |
| weak-distractor | Implausible wrong option gives the answer away |
| duplicate | Near-identical stem/scenario to another item |
| schema | Bad domain, missing studyArea, empty explanations |
| scope | Trivia or undocumented detail rather than a real concept |
Output
List each finding as: id — category — one-line problem — suggested fix. Offer to apply fixes.
When done, run node scripts/build.mjs --check and confirm it passes.