The exam reuses a small set of scenarios and a handful of exact values. Drill these until they're reflexive — they convert "thinking" questions into "recognition" questions.
The 6 recurring scenarios
Roughly four of these appear on any given exam instance. Each tests the same engineering judgement in a different costume.
| Scenario | Domains | What it tests |
|---|---|---|
| Customer Support Agent | 1, 2, 5 | Escalation triggers, structured errors, PreToolUse for refunds, context preservation |
| Code Generation | 3, 5 | CLAUDE.md hierarchy, plan mode, skills, independent review |
| Multi-Agent Research | 1, 2, 5 | Narrow decomposition, parallel Tasks, provenance, gap annotation |
| Developer Productivity | 1, 2, 3 | Grep vs Glob, MCP config, session management |
| CI/CD Pipeline | 3, 4 | -p flag, multi-pass review, independent review, duplicate prevention |
| Structured Extraction | 4, 5 | Nullable fields, self-correction, retry limits, batch API |
Don't memorise scenario-specific answers — the same principles transfer to any costume. When you recognise a scenario, recall its failure modes (above), then apply the 3 rules.
Escalation: 3 valid triggers, 2 traps
Escalate when: (1) the customer explicitly asks for a human — immediate, no "let me try one more thing"; (2) there's a policy gap — no rule for the edge case; (3) no progress after a bounded number of attempts.
Never escalate on: sentiment ("seems frustrated" — frustrated customers often want resolution) or self-reported confidence ("I'm 60% sure" — Claude's raw confidence is uncalibrated). Use explicit categorical criteria.
Conflicting sources
When two sources disagree, annotate both with source and date — never pick one, never average. Averaging invents a number neither source supports. Check publication dates first: a 2022 figure vs a 2024 figure isn't a contradiction, it's a trend.
File locations (exact)
| File | Scope | In VCS | Use |
|---|---|---|---|
~/.claude/CLAUDE.md |
Personal | No | Personal preferences |
.claude/CLAUDE.md |
Team | Yes | Team standards |
src/api/CLAUDE.md |
Directory | Yes | Subsystem rules |
.claude/rules/*.md |
Path glob | Yes | Cross-directory conventions by file type |
.mcp.json |
Team | Yes | Team MCP servers |
~/.claude.json |
Personal | No | Personal MCP servers |
Critical values to memorise
| Item | Value |
|---|---|
| Pass score | 720 / 1000 |
| Questions / duration | 60 MCQ / 120 min |
| Batch API savings / window | 50% / up to 24h, no SLA |
| Max tools per agent | 4–5 |
@import nesting |
5 levels max |
| CI flag | -p / --print |
| Loop done / wants tool | stop_reason = end_turn / tool_use |
"You see X → answer Y" shortcuts
| You see… | Answer is… |
|---|---|
| Financial / compliance / security must-happen | Hook (Pre or PostToolUse) |
| Tool mis-routing | Fix tool descriptions |
| Narrow multi-agent results | Fix coordinator decomposition |
| CI pipeline hangs | Missing -p flag |
| Team standards location | .claude/CLAUDE.md |
| Conflicting statistics | Annotate both, never pick/average |
| Explicit human request | Escalate immediately |
| Absent data + retry loop | Nullable fields instead |
| 18+ tools | Split into scoped agents |
| Self-review | Independent instance |
Context reliability (Domain 5): for values that must survive a long chat, inject a case-facts block verbatim at the top of every prompt — it survives /compact because it's re-injected outside the conversation history.
You've got the patterns. Now drill the 240-question domain quiz →