Solution Architecture

Part of the discovery-phase skill pack · scoping group · reads scope-doc.md (run feature-scoping first if missing) and feeds estimation.

Produces a discovery-grade architecture: enough detail to estimate effort, surface technical risks, and explain choices to the client — but not detailed design. Detailed design happens in delivery, not here. Where the BA isn't a tech lead, this skill is best run pair-mode with an architect; otherwise it produces vague boxes-and-arrows.

Step 1 — Read context + scope

Read discovery-context.md (sections 2. Product / Initiative, 6. Constraints, 7. Tech Context), then scope-doc.md (or mvp-definition.md if MVP track), then risk-assumption-map.md.

If scope-doc.md is missing, recommend feature-scoping first — architecture without a defined scope drifts into "general platform" territory and stops being useful for estimation.

If discovery-context.md is missing, ask inline: "(a) any client-mandated tech constraints (cloud, language, vendor)? (b) integration list — what must this connect to? (c) team's existing stack expertise?" — tag unknowns [ASSUMED].

Step 2 — System context (one diagram, in text)

Write a single system context diagram in text or Mermaid: the system in the middle, actors on one side, external systems on the other. This is the C4 "Level 1" view — no internals.

flowchart LR
  user[End User]
  admin[Admin]
  system[<<system under discovery>>]
  payment[Payment Provider]
  crm[Client CRM]

  user --> system
  admin --> system
  system --> payment
  system --> crm

Explicitly state what's out of scope for this system (e.g., "billing UI is consumed from existing client portal, not built here"). Out-of-scope boundaries are where most estimation surprises hide.

Step 3 — Logical components (3-7)

Decompose the system into 3-7 logical components — single-responsibility boxes. For each:

• Name — <component> (e.g., auth-service, event-ingestion, report-generator)
• Responsibility — one sentence, one job
• Owns data? — yes / no / shared
• Talks to — which other components / external systems
• New or existing? — building from scratch / extending existing client component / adopting a library

3 = too coarse to estimate, 7+ = pre-mature decomposition. If you need 8+, you're in detailed design — wrong skill.

Step 4 — Critical data flows (top 2-3)

Pick the 2-3 most critical or highest-risk flows and walk through them step by step. Examples:

• "User submits an order" — synchronous, must be ACID, 99.9% target
• "Daily reconciliation" — batch, eventually consistent, recoverable
• "Real-time dashboard update" — event-driven, lossy OK

For each flow, note: sync vs async, failure mode, expected latency / volume. These three properties drive infra cost more than feature count and are usually missing from estimations.

Step 5 — Integrations + auth model

For every external system the architecture touches:

If any cell is ?, it's either an open question for the client or a research task. Do not invent rate limits or SLAs — vendor docs are the source of truth.

Auth model gets a separate paragraph: who issues identity, who validates, where session lives, SSO requirements. Auth is the most under-estimated piece in agency proposals.

Step 6 — Non-functional requirements (NFRs)

Concrete, measurable. No "fast", "secure", "scalable" — those are aspirations.

If the client hasn't stated an NFR, state your assumption and tag it [ASSUMED] — don't pretend NFRs don't exist. Unstated NFRs become contract disputes in delivery.

Step 7 — Tech-stack rationale

For each major choice (language, framework, database, queue, deployment target), write 1-2 sentences of rationale anchored on:

1. Constraints — client mandate, existing system, regulatory
2. Team skills — what the agency / client team can actually operate
3. Fit-for-purpose — why this tech matches the workload (not "it's modern")
4. Operational cost — license, hosting, ops effort

Avoid tech religion. If two stacks are equivalent on the first three, pick the one with cheaper ops. Always note the alternatives considered and why rejected — this is what a good architecture review looks for.

Step 8 — Risk surface update

Re-read risk-assumption-map.md. Add any new technical risks surfaced during architecture work (e.g., "third-party SLA below our target", "no existing team experience with X"). For each existing tech risk in the map, note whether the proposed architecture mitigates, accepts, or amplifies it.

If 3+ new HIGH risks appear, the architecture is fragile — loop back, simplify, or escalate to the client.

Step 9 — Pressure test (5 checks)

If 2+ tests fail, the architecture isn't discovery-ready. Loop back rather than feeding shaky inputs into estimation.

Step 10 — Write artifact

Output: ./discovery/solution-architecture.md — see ./template.md.

Append to _log.md: [solution-architecture | YYYY-MM-DD] components: <N>; integrations: <M>; nfr_count: <K>; new_risks: <L>; assumed_count: <P>.

Anti-patterns

• Detailed design. Class diagrams, ORM schemas, API contracts — wrong skill, wrong phase. Discovery architecture is C4 Level 1-2 max.
• Tech-stack religion. "We always use X" without rationale tied to constraints/skills/fit.
• Aspirational NFRs. "Fast, secure, scalable" is not an NFR. Numbers or [ASSUMED].
• Hidden integrations. Every line in the system context diagram has cost. Enumerate or under-estimate.
• Architecture without scope. Producing this skill before feature-scoping produces "platform diagrams" that never match what's being built.
• Over-decomposition. 12 microservices in discovery is fan-fiction. Stay coarse.
• Skipping operational concerns. No deployment / observability / recovery plan = surprise costs in delivery.