The distractors:
- "Repeat the rule verbatim after every paragraph" → bloats the prompt, dilutes attention further
- "Lower temperature to zero" → affects sampling randomness, not rule salience
- "Switch rule to uppercase" → superstition, not a supported mechanism
The answer: Move the critical rule to the beginning or end of the prompt, structurally separated from reference content.
Domain 2 Decision Framework
- Critical rule not being followed → positional attention problem → move to start or end
- Format compliance inconsistent despite instructions → use few-shot exemplars
- CoT applied uniformly → remove from simple extraction tasks, keep on multi-step reasoning
- Cache hit rate near zero despite stable prompt → dynamic values at position zero; fix by moving them after static content
- Mid-collection recall degradation + high cost → retrieve relevant sections, not full corpus
Domain 3 — Integration
Sample Scenario
A Claude-based assistant queries an HR system on behalf of employees. It authenticates using a single service account with organisation-wide read access. The system prompt instructs the model to only return data belonging to the requesting employee.
What the scenario signals: trap 12 — an instruction doing a security layer's job. The model has access to all employee data and is told not to return it. That is not authorization.
The distractors:
- "Service accounts cannot be used with AI systems under compliance frameworks" → not a real rule
- "Add write access for the audit log" → expands the blast radius further
- "Authenticate directly using each employee's password" → a credential anti-pattern
The answer: The core problem is authorization enforced by prompt instruction rather than by the access-control layer. The fix is per-user scoped credentials or pass-through authentication so the model can only retrieve what it is authorized to retrieve.
Domain 3 Decision Framework
- Many AI apps need the same system integration → MCP server
- One deterministic pipeline calls the model once → direct API integration
- Two orgs' agents must coordinate without exposing internal systems → agent-to-agent protocol
- Tool-selection accuracy declining as tools were added → audit/remove unused tools + progressive discovery
- Authorization via system prompt → wrong, always; must be structural access control
- RAG returns clause fragments without context → structure-aware chunking with cross-reference metadata
Domain 4 — Evaluation, Testing & Optimization
Sample Scenario
A marketing-content assistant must be evaluated for tone and brand alignment across thousands of outputs per week. Human review of every output is infeasible.
What the scenario signals: subjective quality at scale → LLM-as-judge methodology.
The distractors:
- "Exact string matching against a library of approved copy" → can't score novel copy
- "Skip tone evaluation, since it is subjective" → abandons a measurable requirement
- "Full human review, accepting the throughput cost" → infeasible as stated
The answer: LLM-as-judge rubric scored against brand guidelines, periodically calibrated against human expert ratings. The calibration step is what makes the judge trustworthy.
Domain 4 Decision Framework
- Output checkable mechanically → code-based eval
- Output requires interpretation → model-based (LLM judge): different model, constrained verdicts, calibrated
- "Good enough to launch?" → task-specific metrics with agreed thresholds, not anecdotes or general benchmarks
- Golden dataset → real queries + constructed edge cases (not synthetic only, not prompt examples)
- Offline eval improved by 6 points → run A/B test on fraction of real traffic before full rollout
- Leading indicator of quality degradation → retrieval relevance scores and "no grounded answer" rates, not support complaints
Domain 5 — Governance, Safety & Risk
Sample Scenario
A document-triage system routes 100% of flagged documents to a human reviewer. The team considers this the cautious approach. The reviewer approves nearly everything after a brief glance.
What the scenario signals: trap 10 — everything reviewed = nothing reviewed. 100% volume with no triage has produced consent fatigue.
The distractors:
- "The review gate is correctly designed; increase the reviewer's throughput" → wrong diagnosis
- "Remove the human review gate since it isn't working" → right observation, wrong fix
- "The system needs a higher confidence threshold before flagging" → partially right framing but misses the routing rule
The answer: Route to human review when low-confidence AND (irreversible OR high-cost). Route by stakes, not by volume. Redesign the gate to let high-confidence, reversible items pass without review.
Domain 5 Decision Framework
- Control named in design doc but no owner and no evidence → named-but-unowned = unenforced (trap 7)
- System relies on system-prompt instruction for data access control → trap 12, always wrong
- 100% routed to human review → trap 10, consent fatigue; fix with confidence + stakes routing rule
- Guardrail that errors silently → fail-open; dangerous for high-stakes data; must decide per-guardrail
- Regulatory obligation with no evidence artifact → a claim, not a control; add owner and evidence
- Indirect prompt injection → screening the user message is insufficient; screen retrieved content too
Domain 6 — Stakeholder Communication & Lifecycle
Sample Scenario
A CFO asks for a 25% cost reduction in the AI platform's monthly spend within 90 days. An architect proposes switching all workloads to the smallest model immediately. The CFO approves.
What the scenario signals: cost optimization without measurement first (trap from M2) + missing tradeoff framing (missing reversal cost from M4).
The correct architect response: Analyse token usage by workload first. Identify dominant cost drivers. Apply targeted levers with quality evaluation at each step. Present to the CFO as: gain (cost reduction path), give-up (engineering time for analysis + risk of quality degradation if downsizing is applied blindly), reversal cost (re-architecture time if a bulk downsize breaks a production workflow and must be reversed).
Domain 6 Decision Framework
- Stakeholder says "seamless" → ask what would break the experience → convert to p95 latency SLA
- Moving to design before all four discovery categories covered → incorrect; missing "must prove" creates compliance gaps
- Tradeoff presented without reversal cost → incomplete; add reversal cost
- Outcome document missing the before-metric → unrecoverable after launch; the before-metric must be captured first
- Governance table doesn't exist before launch → design it in; post-incident governance is incident management
Domain 7 — Developer Productivity & Ops
Sample Scenario
A team of 200 engineers is rolling out Claude. The platform team decides to enable access for all 200 on day one to build momentum quickly.
What the scenario signals: direct batch rollout without champion pattern → produces stalled adoption at basic chat.
The correct architect recommendation: enable one champion per team first; capture their workflow and solved configuration; seed adoption in batches from that shared baseline.
Domain 7 Decision Framework
- Skills distribution: org-provisioned (reach, no rollback) / plugin (versioned) / project Skill (repo-scoped) / API Skill (pinned)
- "We need every team to have this" → org-provisioned (but note: no native rollback)
- "We need to roll back if this breaks the finance workflow" → plugin (versioned, rollback)
- Spend posture missing → configure model defaults, allowlists, caps before the first bill
- AI-generated code reviewed by the same session that wrote it → session anchoring trap; use fresh session or human
- Human understanding dimension in verification checklist → never automated; always requires a human to verify they actually understand
Scenario-Matching Practice
The exam includes questions where you match five scenarios to four options (one option used twice, or one option not used). The distinguishing logic:
Architecture pattern matching:
- One transformation + supplied context → single augmented call
- Same steps every time → fixed workflow
- Emergent path → autonomous agent
- Different specialisations + parallel → multi-agent
Integration mechanism matching:
- Standardised reuse across many apps → MCP server
- Deterministic owned pipeline → direct API integration
- Cross-org coordination without internal exposure → agent-to-agent protocol
Skills distribution matching:
- Entire org, stable capability → org-provisioned
- Group-targeted, needs rollback → plugin
- Repo-scoped procedure → project Skill
- Programmatic, version-pinned → API Skill
Final Pre-Exam Checks
Before sitting the exam, confirm you can answer each of these without looking anything up:
- What are the three feasibility verdicts?
- Which layer of the safety stack catches indirect prompt injection?
- What is the HITL routing rule, stated exactly?
- What three fields does the compliance control register require?
- What is the reversal cost of the architecture you're recommending in a tradeoff?
- Which reliability control belongs at the service boundary (not the API layer)?
- What is the one dimension of the verification checklist that can never be automated?
- Which distribution mechanism supports native version rollback?
- What is the difference between shadow testing and canary rollout?
- Why is an uncalibrated LLM judge worse than no automated grade?
If any of these take more than 10 seconds, revisit the corresponding module lesson before proceeding to the exam.