Question Anatomy
Remember
- The last sentence decides the answer.
- Same story + different last sentence = different correct answer.
- Hard questions hide the tie-breaker in one small detail of the story.
Structure of every question
flowchart TD
A["Long scenario story<br/>(context, symptoms, constraints)"] --> B["THE ASK<br/>(last sentence)"]
B --> C["4 options<br/>2 easy to kill, 2 plausible"]The ask — read it first
| The ask says… | It wants… | Trap options are… |
|---|---|---|
| "primary reason" / "most likely cause" | A diagnosis | Valid fixes (but not causes) |
| "most effective fix" | The change with biggest direct impact | Smaller partial fixes |
| "best first step" | The cheapest, safest first action | Good actions done too early |
| "minimize cost" | The cheap option | Better-quality but pricier options |
| "maximize quality / reliability" | The robust option | Cheaper shortcuts |
| "guarantee / cannot be left to discretion" | Code or hooks, not prompts | "Add prompt instructions" |
Difficulty tiers
- Easy — only one option makes sense.
- Medium — distractors add one extra detail that does not fit the story.
- Hard — two options are nearly identical. The tie-breaker is a small detail: a number, a policy, a stakeholder need. Re-read the story hunting for it.
Risk words that mark WRONG options
Multiple independent test-takers report these words usually sit in the wrong answer:
- "automatically" (for risky actions)
- "unrestricted" / "full access"
- "always" / "never"
- "without review" / "without confirmation"
Anthropic's style = least privilege, human oversight, right-sized autonomy.
Common distractor styles
- ❌ Re-doing work instead of preserving it (re-search sources, re-read all files).
- ❌ Passing raw dumps instead of structured summaries.
- ❌ Fixing symptoms downstream instead of the source.
- ❌ "Just add more prompt instructions" for things that must be guaranteed.
- ❌ Extreme options: "skip summarization entirely", "retry forever", "always escalate".
Next: Study Plan