The exam presents 4 of these 6 scenarios at random, each framing several questions. Know the setup, the tools in play, the domains each pulls from, and the traps each scenario tends to bait.
Agent SDK agent handling returns, billing disputes, account issues via MCP tools get_customer, lookup_order, process_refund, escalate_to_human. Target: 80%+ first-contact resolution while knowing when to escalate.
Favored questions: prerequisite gates (verify before refund), interception hooks (refund caps), tool-description confusion (get_customer vs lookup_order), structured errors, escalation calibration, case-facts blocks, multiple-match disambiguation, multi-concern decomposition.
Traps: using a prompt where a guarantee is needed; escalating on sentiment/self-confidence; picking a customer heuristically on multiple matches; over-engineering a classifier when descriptions/criteria suffice.
Claude Code for generation, refactoring, debugging, docs. Custom slash commands, CLAUDE.md config, plan vs direct execution.
Favored questions: where to put a shared /review command; user vs project config; glob rules for scattered files; plan mode vs direct; interview pattern; interacting-vs-independent fixes.
Traps: user-scope config for team-wide needs; per-directory CLAUDE.md for files spread across the tree; fabricated .claude/config.json commands array; "start direct, switch later" when complexity is already stated.
Coordinator delegates to subagents: web search, document analysis, synthesis, report generation. Produces comprehensive, cited reports.
Favored questions: narrow-decomposition root cause; context isolation of subagents; parallel Task calls; scoped verify_fact tool; structured error propagation on timeout; claim-source provenance; conflicting-source annotation; coverage-gap iterative refinement.
Traps: blaming a downstream agent for an upstream decomposition gap; giving the synthesis agent the full toolset; silently suppressing a subagent failure or killing the whole workflow; averaging/dropping conflicting figures instead of annotating.
Agent SDK tools to explore unfamiliar codebases, understand legacy systems, generate boilerplate, automate tasks. Uses built-ins (Read, Write, Bash, Grep, Glob) + MCP servers.
Favored questions: Grep vs Glob; Edit non-unique → Read+Write; incremental exploration; Explore subagent for context; scratchpad files; MCP resources for catalogs; enhancing MCP tool descriptions over built-in Grep; fork_session for divergent approaches; crash-recovery manifests.
Traps: reading all files upfront; bigger window "fixes" context degradation; preferring a weak built-in because the MCP description is thin.
Claude Code in CI/CD: automated code review, test generation, PR feedback. Prompts must give actionable feedback and minimize false positives.
Favored questions: -p to avoid hangs; --output-format json+--json-schema; explicit review criteria (not "be conservative"); disabling high-false-positive categories; independent review instance; report-only-new on re-runs; batch vs sync for blocking pre-merge; per-file + integration passes.
Traps: fabricated flags (--batch, CLAUDE_HEADLESS); self-review of generated code; bigger window for a 14-file review; batching a blocking pre-merge check; confidence-filtering instead of categorical criteria.
Extracts info from unstructured documents, validates with JSON schemas, high accuracy, graceful edge-case handling, integrates downstream.
Favored questions: tool_use + JSON schema; syntax vs semantic errors; nullable fields to stop fabrication; enum other/unclear; retry-with-error-feedback and its limits; calculated-vs-stated validation; Batches API fit + custom_id; few-shot for varied structures; segment accuracy by type/field; calibrated field-level confidence.
Traps: assuming a schema catches semantic errors; keeping a field required and "forbidding" fabrication; retrying when info is absent from source; automating on aggregate accuracy; trusting raw self-reported confidence.