Domain 1: Solution Design & Architecture

Cheat sheet for Solution Design: framing, e2e stages, patterns, multi-agent, decomposition, value pillars.

Business → Claude solutions (1.1)

Frame before stack. Map goals, constraints, owners, I/O, non-goals, and success metrics. Claude for judgment-heavy language; deterministic ops stay in code.

Refuse vague "AI on X" until outcomes are explicit
Non-goals prevent CRM-wide scope sprawl
Demo delight ≠ production success criteria
Model/RAG/prompt choices come after framing

End-to-end architecture (1.2)

Stages: input → processing → output → feedback
Failures, retries, escalations re-enter the loop
HITL and monitoring are first-class when risk requires them
Feedback must feed evals/prompts/retrieval — not only a thank-you UI
Diagrams that stop at "model responds" fail

Architectural patterns (1.3)

Workflow = fixed gates. Agentic = open tool plans. Augmented LLM = single call + retrieval/light tools. Prefer the simplest pattern that meets control needs; hybrids OK by stage.

Fixed sequential claims/compliance → workflow
Open exploration under a toolkit → agentic
Bounded corpus Q&A → augmented LLM
Not multi-agent by default

Multi-agent orchestration (1.4)

Coordinator owns decomp, routing, aggregation, visibility
Specialists narrowly scoped; explicit handoff contracts
No inherited/shared memory across subagents
Hub-and-spoke beats peer mesh for audit
Justify multi-agent by isolation or parallel specialization

Decomposition (1.5)

Split by dependency, uncertainty, verification boundaries
Unknown structure → adaptive decomposition
Re-merge with explicit synthesis / coverage criteria
Thin categories after good specialists → bad coordinator split
Avoid whole-problem monolithic dumps when units are separable

Business value pillars (1.6)

Pillars: efficiency, transformation, productivity, cost, performance SLAs. Name the pillar, measure the trade-off, keep SLAs in the architecture decision.

Cost + hard p95 above accuracy floor → smaller/faster model
Do not mislabel low-volume automation as transformation
Document rejected alternatives when pillars conflict
Dropping monitoring to "save cost" is not aligned