Worked example — auditing a production marketing site

This is a real run of scripts/design-audit.mjs against a live single-page marketing site (a financial-services firm's homepage), captured to show what the heuristic layer of the stack produces before you wire it into your own project.

How it was produced

bash

# serve the site, then audit it over HTTP (not file://, so asset/CORS noise is realistic)
python3 -m http.server 8099          # in the site directory
node scripts/design-audit.mjs --url http://localhost:8099/ --out examples/juniper-audit

What's here

report.md — ranked findings + console health
report.json — the same, machine-readable (what CI consumes)
screenshots/ — full-page capture at all six viewport tiers (360 → 1920 px)

Reading the results

The site is genuinely well-built (design tokens, self-hosted fonts, prefers-reduced-motion, scroll-reveal). The audit still surfaced three actionable items — exactly the kind a human misses on a quick look:

Finding	Severity	Why it matters
2–3 sampled interactive elements show no visible focus indicator	High	Keyboard users can't see where they are — WCAG 2.4.7
14 tap targets < 44×44 px on mobile	Medium	Small touch targets → mis-taps on phones
~1 text node below AA contrast (approximate)	Medium	Lead to verify by hand; contrast is estimated from the nearest opaque background

The single console error is a 404 for one asset — worth chasing, but not a layout defect.

Two things this example teaches about the tooling

Serve over HTTP, not file://. Opening via file:// produces a wall of font-CORS errors that are pure artifacts (browsers block cross-origin font fetches from null origin). The numbers above are from a real HTTP serve, so the console is clean.
Full-page capture + aggressive scroll-reveal don't fully mix. The script auto-scrolls to trigger IntersectionObserver reveals before shooting, which recovers most content — but a page that reveals sections only on fine-grained scroll can still show a blank band in a single full-page screenshot. That's both a tooling limit and a real robustness signal (the same content is invisible to some crawlers, print, and any non-interactive capture). For those pages, the interactive /design-review subagent — which scrolls naturally, waits, and interacts — is the more reliable check. Heuristics catch the mechanical defects; the subagent judges the experience.

The point

This is the fast, automatic half of the stack (also runs in CI). The full review — interaction flows, visual taste, edge cases — is the design-review subagent invoked with /design-review <url>. Use both: heuristics to gate merges, the subagent to raise the bar.