Episode Eleven completes Professional by turning the architecture into decisions, delivery, ownership, and continuous improvement. By the end, you should be able to document requirements and trade-offs, hand the system to operations, manage lifecycle change, enable developers safely, and connect the Professional responsibilities into one production view. Now carry the complete architecture through stakeholder agreement, implementation, operations, and change. Stakeholder communication is worth fourteen percent of the exam, while developer productivity and operational enablement add seven percent. That combined weight means a technically elegant design can still be the wrong answer if nobody can approve, build, operate, or improve it. Structured discovery begins with the people who perform, govern, support, and experience the process.

Interview users, product owners, domain experts, security, privacy, legal, data owners, operations, support, procurement, and executive sponsors as relevant. Ask for real cases, exceptions, current metrics, workarounds, failure consequences, and decision rights. Observe the actual workflow because stated process and lived process often differ. How should requirements be documented? Separate outcomes, functional needs, non-functional constraints, assumptions, dependencies, risks, and acceptance measures. Rank them so the architecture can resolve trade-offs openly. An executive may prioritise rapid rollout, while operations require diagnosability and legal requires a defined human decision point. The architect should show how options affect each priority rather than hiding the conflict in technical detail.

A useful decision record states context, options, decision, rationale, consequences, owner, date, and review trigger. For example, Harbour Resolution may choose assisted drafting before automatic action because it produces measurable value while preserving officer accountability during evidence gathering. The record should also say what evidence could justify greater automation later. What diagrams help? A context diagram, component view, data-flow diagram, trust boundaries, sequence for critical journeys, deployment view, and operational ownership map. However, use the minimum set that answers stakeholder questions. A data-flow diagram should show sensitive data entering prompts, retrieval, tools, logs, caches, evaluations, and human interfaces. A sequence diagram should show identity, model requests, tool use, validation, approval, and side effects for a high-risk journey.

An architecture document should identify systems of record, interfaces, failure modes, service levels, security controls, and unresolved decisions. Implementation guidance should be concrete enough that delivery teams do not reinterpret every boundary. Include tool contracts, prompt and schema ownership, versioning, evaluation gates, logging requirements, environment separation, and release controls. Now move through the lifecycle. Discovery establishes value, users, constraints, and baseline. Design selects patterns, boundaries, controls, and success criteria. Prototype tests the riskiest assumptions rather than polishing the easiest demonstration. Pilot runs a bounded real workflow with representative users and monitored fallback. Production hardening adds identity, scale, observability, support, recovery, governance, and tested operations. Handoff establishes accountable owners, runbooks, training, service levels, and escalation paths.

Monitoring and iteration use real evidence to update prompts, retrieval, controls, and process. Retirement removes access, secrets, indexes, data, schedules, and unsupported dependencies according to policy. What should a handoff package contain? Architecture decisions, deployment and rollback steps, configuration inventory, data flows, evaluation results, dashboards, alerts, runbooks, owners, known limitations, and incident procedures. Also include a change process for models, prompts, Skills, tools, schemas, and source corpora. A model update is a production change that may require regression evaluation even when the A P I remains compatible. A policy-source refresh is also a production change because it can alter retrieval and answers. Stakeholder feedback loops need design.

Users should be able to correct outputs and identify missing evidence without submitting vague thumbs-down signals only. Domain owners should review patterns of failure and approve source or policy changes. Engineering should receive traceable defects with versions and reproducible inputs. Governance bodies should see risk, control performance, incidents, exceptions, and benefit measures at an appropriate level. Executives need outcome and exposure, not a wall of token statistics. Now developer enablement. A team rollout of Claude Code should provide approved installation and identity paths, repository guidance, managed settings where needed, safe defaults, reusable Skills, example prompts, and support. Create golden workflows for common tasks such as codebase orientation, test creation, migration review, secure pull-request review, and documentation.

Keep shared configuration version controlled and reviewable. Use hooks and C I for mandatory checks rather than relying on every developer to remember a sentence. Measure adoption together with accepted change quality, review effort, defect rates, cycle time, cost, and developer experience. High usage alone does not prove productivity. Provide a channel for reporting unsafe behaviour, tool errors, and workflow friction. Teach developers how to scope tasks, supply acceptance criteria, review diffs, protect secrets, and isolate parallel work. Operational enablement also means debugging the A I system as a distributed application. Support staff need trace identifiers and stage-level telemetry rather than screenshots of a surprising answer.

Runbooks should distinguish provider errors, rate limits, retrieval failures, tool failures, validation failures, permission denials, and human-queue delays. Now the exam-day method. Read the scenario before optimising the answer. Identify the stated goal, constraints, risk, and stage of the lifecycle. Separate model behaviour from application guarantees. Prefer root-cause controls over cosmetic safeguards. Remove unused privilege before adding prompts or monitoring. Preserve evidence, state, and failure semantics. Choose architecture proportional to task variability and consequence. Measure the end-to-end outcome, not just the model call. For every proposed control, ask who enforces it and how you would prove it operated. For every proposed optimisation, ask what assumption it introduces and how you will detect regression.

For every human review, ask what decision the human makes and what evidence they receive. For every integration, ask whose identity reaches the service and where authorisation is enforced. For every R A G answer, ask which source version was retrieved and whether it was applicable. For every agent, ask whether a workflow or narrower assistant would achieve the same value more safely. The series is complete when you can connect business value, architecture, model behaviour, integration, evaluation, governance, delivery, and operations without losing the causal chain. For source navigation, combine both exam guides with the Claude Code feature, routine, error, goal, and multi-agent operations references, then refresh all living details before sitting the exam.

The best final preparation is to build one end-to-end solution with retrieval, tools, structured outputs, evaluation, observability, and a controlled human decision. Then explain its architecture aloud without relying on product slogans. You are ready when you can describe how it succeeds, how it fails, how it is governed, and how evidence changes your next design decision.