Skip to main content

Tutorial: Run a Frontend Architecture Review Program

Why this tutorial matters

Architecture work fails when it stays inside one architect's head. Teams need a review program that turns decisions into visible artifacts, reusable checklists, scorecards, and follow-up actions without becoming bureaucracy.

This tutorial builds a practical Frontend Architecture Review Program for a mid-sized product organization.

What you will build

ArtifactPurpose
intake templatecapture proposed change and risk
ADR templaterecord important decisions
RFC processreview cross-team changes
scorecardstrack health and ownership
review checklistsperformance/security/accessibility/design system
exception registermanage deviations with expiry
postmortem loopturn incidents into architecture controls

Milestone 1: Define review scope

Not every change needs architecture review.

Review required when a change:

  • affects shared rendering/data architecture
  • adds a third-party script
  • changes auth/session behavior
  • introduces new generated UI or AI tool action
  • changes design-system primitives
  • crosses team boundaries
  • affects Core Web Vitals on critical routes
  • creates a long-lived exception

No review required for:

  • local feature UI using established patterns
  • copy/content changes inside approved templates
  • small bug fixes with no contract change

Acceptance criteria:

  • teams know when review is required
  • review scope is risk-based, not title-based

Milestone 2: Create intake

Use a one-page intake:

# Architecture Review Intake

## Change
What are we changing?

## User impact
Who is affected and how?

## Risk areas
- Performance:
- Accessibility:
- Security:
- Reliability:
- Data/rendering:
- Design system:

## Reversibility
How do we roll back or degrade?

## Decision needed
What question should reviewers answer?

Acceptance criteria:

  • reviewers can understand the risk in five minutes
  • teams are not forced to write a long document for low-risk changes

Milestone 3: ADR workflow

ADR required for decisions that are hard to reverse or teach.

ADR sections:

  • context
  • decision
  • alternatives
  • tradeoffs
  • consequences
  • owner
  • review date
  • links to metrics/incidents

Rules:

  • ADRs are short
  • rejected alternatives are included
  • review date is explicit
  • ADRs link to implementation PRs

Milestone 4: RFC workflow

Use RFCs for cross-team or platform-level changes.

RFC phases:

  1. draft
  2. review
  3. decision
  4. implementation
  5. adoption
  6. follow-up

Decision record:

FieldPurpose
statusproposed, accepted, rejected, superseded
owneraccountable person/team
affected teamsadoption planning
migration planprevents orphaned old paths
success metricsmakes decision testable
rollbackprotects production

Milestone 5: Scorecards

Create route/system scorecards.

ScorecardExample metrics
performanceLCP, INP, CLS, JS KB, request count
securityCSP status, third-party owners, dependency risk
accessibilityautomated violations, manual test coverage, known defects
reliabilityerror rate, degraded-mode coverage, incident count
design systemcomponent adoption, token usage, overrides
AI/GenUIinvalid schema rate, approval rate, tool failures, eval pass rate

Scorecards should show ownership and trend, not just numbers.

Milestone 6: Exception register

Every exception needs:

type ArchitectureException = {
id: string;
rule: string;
reason: string;
owner: string;
createdAt: string;
expiresAt: string;
mitigation: string;
};

Examples:

  • route exceeds JS budget until migration completes
  • third-party script allowed for campaign period
  • CSP temporarily allows one external domain
  • design-system escape hatch used for partner launch

Acceptance criteria:

  • exceptions expire
  • owners are visible
  • expired exceptions are reviewed

Milestone 7: Review meeting format

Keep meetings short:

  1. decision needed
  2. risk summary
  3. alternatives
  4. tradeoffs
  5. decision or required follow-up

Do not use review meetings for live document reading. Documents must be readable before the meeting.

Milestone 8: Incident-to-architecture loop

After an incident, ask:

  • which architecture boundary failed?
  • what signal should have detected it earlier?
  • what default would prevent recurrence?
  • what checklist needs a new item?
  • what test/eval/CI gate should be added?
  • what ADR/RFC should be updated?

This converts incidents into institutional learning instead of blame.

Program health metrics

Track:

  • review cycle time
  • number of expired exceptions
  • recurring issue categories
  • ADRs without review date
  • scorecard coverage
  • incident follow-up completion
  • teams using templates without architect handholding

The goal is not more reviews. The goal is better decisions with less repeated explanation.

Review checklist library

Create short checklists, not giant forms.

Performance:

  • route budget changed?
  • LCP element affected?
  • new JavaScript or third-party domain?
  • new client-side experiment/personalization?
  • rollback path?

Security:

  • new script, iframe, rich text, generated UI, or auth boundary?
  • CSP/security headers affected?
  • sensitive data in browser/cache/logs?
  • server validation present?

Accessibility:

  • keyboard path tested?
  • focus behavior changed?
  • async status announced?
  • error messages associated?
  • color/contrast checked?

Design system:

  • uses existing primitives?
  • new component justified?
  • tokens used correctly?
  • escape hatch documented?

AI/GenUI:

  • model can only propose allowed outputs?
  • generated UI schema validated?
  • retrieval permissions enforced?
  • tools classified by side effect?
  • evals updated?

Example review: edge-routed GenUI landing page

Intake summary:

  • AI generates landing page variants from ad intent.
  • Humans approve variants.
  • Edge middleware routes to static variant path.
  • Small form island hydrates.

Review findings:

  • performance review requires CDN cache key proof
  • security review requires no user data in cache key
  • accessibility review requires generated variant templates pass checks
  • design-system review requires approved section components
  • AI review requires source/copy policy and human approval record

Decision:

  • approved for one route with three variants
  • variant expansion requires separate review
  • rollback flag must disable edge rewrite
  • metrics: LCP, INP, conversion, invalid variant, cache hit ratio

Operating cadence

CadenceActivity
weeklyreview high-risk intakes
biweeklyreview scorecard regressions
monthlyexpired exceptions review
quarterlyarchitecture standards update
after incidentspostmortem-to-checklist update

Keep cadence predictable. Ad hoc architecture review becomes either a bottleneck or an afterthought.

Capstone exercise

Run the program on three proposals:

  1. Add a new analytics vendor.
  2. Introduce generated UI into a customer workflow.
  3. Move a high-traffic route from SSR to edge-routed static variants.

For each, produce:

  • intake
  • ADR or RFC
  • risk checklist
  • scorecard entry
  • exception if needed
  • follow-up review date

Source lens

This tutorial is synthesized from the book's ADR templates, architecture review checklists, NFR templates, source-derived review playbooks, incident postmortem templates, and governance chapters.