Exam-day strategy
With 60 items in 120 minutes you have two minutes per item, and scenario stems are long. Read the scenario once carefully, then answer its items against that picture.
Pacing
- Checkpoint Items done Time used
- End of scenario 1 About 15 30 minutes
- End of scenario 2 About 30 60 minutes
- End of scenario 3 About 45 90 minutes
- End of scenario 4 60 110 minutes, leaving 10 to review flags
Flag and move on after three minutes. Unanswered items score nothing, so answer every one before time runs out. Multiple-response items tell you how many to pick; pick exactly that many.
The four distractor patterns
The official samples reuse four kinds of wrong answer. Learn to spot them and most items come down to two options.
The probabilistic fix for a deterministic need. "Strengthen the prompt" or "add few-shot examples" when the requirement is a guarantee. If money, identity or compliance must never fail, choose the hook or gate.
The over-engineered fix. A trained classifier, keyword router, new model tier or extra infrastructure, offered before the cheap root-cause fix (a better description, explicit criteria) has been tried. Watch for the words "first step" and "most effective".
The wrong component. The symptom shows up downstream, but the logs in the stem point upstream, such as a narrow decomposition by the coordinator. Re-read the evidence the question gives you.
The invented feature. Plausible-sounding flags, files or settings that do not exist, such as -- batch, CLAUDE_HEADLESS or a commands array in a config file.
Ten principles that answer most questions
Guarantees go in code (hooks, gates); preferences go in prompts.
Tool descriptions are the first fix for tool selection.
Subagents know only what you put in their prompt.
Route everything through the coordinator.
Fewer, narrower tools per agent.
Errors carry category, retryability, what was tried and partial results; empty is not the same as failed.
Explicit criteria and few-shot examples beat "be careful" and confidence scores. Tool use with schemas guarantees syntax, never semantics; make absent data nullable. Batch only what can wait 24 hours.
A fresh instance reviews better than the author; split big reviews per file plus integration.
Exam-day strategy