Wrong options are not random — they're drawn from a recurring set of anti-patterns. Memorise the pattern → correct fix mapping and most questions collapse to "which option is not on this list."
The full bank
| # | Anti-pattern (wrong) | Correct fix |
|---|---|---|
| 1 | Check text content to detect loop end | Use stop_reason == "end_turn" |
| 2 | Iteration cap as the primary stop | stop_reason primary; cap = safety net |
| 3 | Parse natural language for completion | Use stop_reason, never parse text |
| 4 | Prompt instruction for critical ordering | PreToolUse hook — deterministic |
| 5 | Subagents communicate directly | All communication through the coordinator |
| 6 | Assume subagents inherit context | Pass all context explicitly |
| 7 | One Task per turn for "parallel" | Multiple Task calls in one response |
| 8 | --resume when tool results are stale |
Fresh start + injected summary |
| 9 | Generic error response | Structured: isError, errorCategory, isRetryable |
| 10 | Timeout → return empty success | Return isError: true with partialResults |
| 11 | Few-shot to fix tool mis-routing | Rewrite the tool descriptions |
| 12 | 18+ tools in one agent | Split into 4–5 tools per scoped agent |
| 13 | Team standards in ~/.claude |
.claude/CLAUDE.md (project, in VCS) |
| 14 | Directory CLAUDE.md for cross-dir rules |
.claude/rules/ with glob frontmatter |
| 15 | Start direct when complexity is stated | Plan mode first when complexity is known |
| 16 | CI without -p |
Always -p / --print for non-interactive |
| 17 | Same-session self-review | Independent instance, zero prior context |
| 18 | Required fields for absent data | Nullable + remove from required |
| 19 | Batch API for real-time needs | Sync when someone waits or tool loops |
| 20 | Retry on absent data | Return null; retry for format errors only |
| 21 | Sentiment/confidence-based escalation | Explicit categorical criteria |
| 22 | Pick or average conflicting stats | Annotate both with source + date |
Grouped by the 3 rules
Rule 1 violations (deterministic > probabilistic): #1–4, #13, #15, #21. Any "add a prompt instruction" or "parse the text" for a high-stakes or must-happen step.
Rule 2 violations (root cause > symptom): #11, #12, #14, #17. Patching routing/retry/sessions instead of fixing the description, tool count, rule scope, or review independence.
Rule 3 violations (simplest correct intervention): #7, #8, #19, #20. Adding complexity, the wrong API, or extra retries where a minimal structural fix exists.
The three that show up most
- #10 — Silent suppression. A tool catches a timeout and returns
{results: [], success: true}. The coordinator believes the search succeeded and found nothing → tells the customer "no orders" during an outage. Always surface a structured error withisRetryable. - #17 — Same-session self-review. The session that wrote the code is biased toward justifying it. Use an independent review instance with zero prior context.
- #16 — CI without
-p. The pipeline hangs forever waiting for interactive input.-p/--printis the single most-tested CI flag.
Drill habit: for each wrong option in the practice quiz, name which anti-pattern number it is. If you can label all three distractors, you understand the question.
Next: Scenarios & rapid recall, then drill the domain quiz.