The Six Scenario Maps
Scenario 1: Agentic Coding Assistant
Context: Claude Code agent helping developers write, test, and review code.
Domain emphasis:
- D1 (heavy): Agent loop, plan mode decision, task decomposition
- D3 (heavy): CLAUDE.md hierarchy, slash commands, CI/CD integration
- D4 (moderate): Few-shot for code style, structured output for code review
Most likely trap:
Using prompts for code quality enforcement when hooks/CI are the answer. “Always format with prettier” → CI rule, not prompt.
Must know cold:
- When plan mode triggers (45+ files, multiple approaches)
- The -p flag requirement for CI/CD usage
- How CLAUDE.md hierarchy works (project vs user vs directory)
Scenario 2: Document Processing Pipeline
Context: Extracting structured data from invoices, contracts, or other documents.
Domain emphasis:
- D4 (heavy): JSON schemas, tool_choice, few-shot for format, retry patterns
- D5 (heavy): Validation layers, confidence scoring, batch API, human review routing
- D2 (moderate): Tool descriptions for extraction schemas, error categories
Most likely trap:
Using tool_choice: ‘auto’ when document type is unknown (should be ‘any’), or using Batch API for real-time pipelines.
Must know cold:
- auto vs any vs forced tool_choice and when each applies
- Schema eliminates syntax errors but not semantic errors
- Batch API = up to 24 hours, never for latency-sensitive work
Scenario 3: Customer Support Agent
Context: Multi-turn customer service agent with refund processing, identity verification, escalation.
Domain emphasis:
- D1 (heavy): Loop control, hook enforcement for compliance, human-in-the-loop
- D2 (heavy): Tool scoping, error categories, scoped access by agent role
- D5 (moderate): Confidence routing for unusual cases, structured error propagation
Most likely trap:
Prompt-based identity verification instead of PreToolUse hooks. “Always verify identity before refunds” in prompt = probabilistic. Hook = deterministic.
Must know cold:
- Programmatic enforcement vs prompt guidance (the #1 most-tested concept)
- isRetryable in error responses
- Human-in-the-loop triggers: scope exceeded, irreversible high-blast, ambiguous
Scenario 4: Research and Synthesis
Context: Multi-agent research coordinator with parallel web search, synthesis, source attribution.
Domain emphasis:
- D1 (heavy): Coordinator-subagent, parallel execution, hub-and-spoke
- D5 (heavy): Source attribution, conflict detection, partial failure, confidence in synthesis
Most likely trap:
Coordinator trap — decomposing a broad topic into subtasks that miss entire categories. The agent’s subtasks succeed but the synthesis is incomplete.
Must know cold:
- Coordinator trap and the scope verification step
- Structured error propagation from failed subagents to coordinator
- Conflict detection: present both conflicting values, not one
Scenario 5: Automated Code Review
Context: Claude Code in CI/CD reviewing pull requests for security, coverage, architecture.
Domain emphasis:
- D3 (heavy): CI/CD integration, headless mode, -p flag, review pipelines
- D4 (heavy): Multi-instance review, per-file vs all-at-once, confidence
- D5 (moderate): Confidence thresholds for blocking vs warning
Most likely trap:
Self-review (same session) vs independent review (separate call). Reviewer must not have generation context.
Must know cold:
- The -p flag is REQUIRED for CI/CD (without it, pipeline hangs)
- Independent review = separate API call, no generation history
- Specialized review stages > generalist review
Context: Processing thousands of documents — invoices, forms, records — for downstream processing.
Domain emphasis:
- D4 (heavy): Schemas, few-shot, retry patterns, validation
- D5 (heavy): Batch API, context efficiency, stratified sampling
- D2 (moderate): Tool schemas for extraction, error handling
Most likely trap:
Using Batch API for blocking pipelines, or missing the two-layer validation (schema catches syntax, semantic validation catches wrong-but-valid).
Must know cold:
- Batch API: 50% cost, up to 24h, no SLA — never for user-blocking
- Two validation layers: schema = syntax, semantic = business rules
- Stratified sampling (not random) for quality validation
Key Takeaways
- Know all 6 scenarios — 4 are selected randomly, you can’t predict which
- Scenario 3 (Customer Support) + Scenario 1 (Coding) most heavily test D1
- Scenario 2 (Documents) + Scenario 6 (Extraction) most heavily test D4
- Scenario 4 (Research) most heavily tests D5 source attribution
- Scenario 5 (Code Review) is where D3 CI/CD and D4 multi-instance intersect