Business problem framing
Translating a vague stakeholder ask into goals, constraints, owners, inputs/outputs, non-goals, and measurable success criteria before choosing models or integrations.
Exam context: Refuse "AI on X" until outcomes and Claude-vs-code fit are explicit.
See also: 1.1 Business → Claude solutions
Claude fit vs deterministic software
Use Claude for judgment-heavy language tasks (nuance, drafting, exception explanation); keep sorts, checksums, fixed field updates, and hard rule checks in code.
Exam context: SQL sorts and Redis counters are not Claude-shaped jobs.
See also: 1.1 Business → Claude solutions
Non-goals
Explicit statements of what the solution will not cover in this phase — critical to prevent scope sprawl after a vague AI ask.
Exam context: Missing non-goals are a framing failure, not a prompt failure.
See also: 1.1 Business → Claude solutions
End-to-end stages
The full Claude architecture path: input (intake) → processing → output (delivery) → feedback (outcomes that improve the system).
Exam context: Diagrams that stop at "model responds" are incomplete.
See also: 1.2 End-to-end architectures
Feedback loop
A path that captures corrections, failures, escalations, and outcome signals and feeds evals, prompts, retrieval, or ops — not only a thank-you UI.
Exam context: Missing feedback in an otherwise complete intake→output path is a classic trap.
See also: 1.2 End-to-end architectures
HITL as architecture stage
Human-in-the-loop review treated as an explicit processing/output stage when risk requires it, with handoff context — not a slide-deck afterthought.
Exam context: HITL belongs on the diagram when irreversible or high-risk outputs are in scope.
See also: 1.2 End-to-end architectures
Workflow pattern
Architectural pattern for fixed sequential steps with explicit gates and deterministic checks between stages.
Exam context: Fixed claims/compliance gates favor workflow over open agentic loops.
See also: 1.3 Architectural patterns
Agentic pattern
Architectural pattern for open-ended tool use and adaptive plans under a bounded toolkit, when the next step is not fully known upfront.
Exam context: Match to unpredictability and tool exploration — not to buzzwords.
See also: 1.3 Architectural patterns
Augmented LLM
Single-call (or short) generation pattern enhanced with retrieval or light tools — preferred when the task is mostly answer generation over a bounded corpus.
Exam context: Prefer augmented LLM over multi-agent when simplicity meets the need.
See also: 1.3 Architectural patterns
Coordinator (hub-and-spoke)
The orchestrating agent that owns decomposition, routing, aggregation, and visibility; specialists communicate through it with explicit handoffs.
Exam context: Peer-to-peer specialist chatter that hides the coordinator fails exam items.
See also: 1.4 Multi-agent orchestration
Explicit context handoff
Passing the facts, constraints, and artifacts a specialist needs in a structured contract — subagents do not inherit coordinator history or shared memory.
Exam context: Assumed inherited memory is a multi-agent distractor.
See also: 1.4 Multi-agent orchestration
Adaptive decomposition
Splitting work as structure is discovered (e.g., unknown codebases, open research) rather than locking a fixed pipeline before exploration.
Exam context: Unknown legacy + open-ended goals favor adaptive splits.
See also: 1.5 Decomposition techniques
Coverage / synthesis criteria
Explicit checklist of categories or acceptance tests used when re-merging specialist outputs so thin or missing areas are caught.
Exam context: Thin coverage after good specialist work usually means a bad coordinator split.
See also: 1.5 Decomposition techniques
Business value pillars
Named value dimensions for architecture alignment: efficiency, transformation, productivity, cost, and performance SLAs.
Exam context: Answers should name the pillar and the measurable trade-off.
See also: 1.6 Business value pillars
Pillar–trade-off alignment
Documenting which pillar a design choice serves, what was rejected, and how success will be measured — especially when cost/SLA conflict with marginal accuracy.
Exam context: Ignore-SLA accuracy wins and mislabeled "transformation" fail.
See also: 1.6 Business value pillars