The 60-second check before you approve any artboard — tested blind: the same design lost 0–3, then won 3–0. Claude Code's new /design flow ships whatever you approve at ~96% measured fidelity, including the defects you approve. On Sep 1, 2026 we ran the test nobody in the /design hype wave ran — a pre-registered 3-brief × 2-arm battery (byte-identical prompts; a greenfield barbershop page, a redesign of our real production Next.js homepage, and a SaaS dashboard), 9 blind verdicts from 3 judges per brief with sealed random label mappings. The /design page lost the greenfield matchup 0–3 because of two flaws sitting IN the artboard we approved: a literal [ADD PHONE] placeholder that shipped verbatim into the footer, and a .topnav a selector silently overriding the Book-a-Chair button's label color at higher specificity. Two edits — sixty seconds of auditing — and a fresh blind panel flipped the identical matchup to a 3–0 clear win. Meanwhile /design swept 3–0 unaided on the one case nobody demos: redesigning a real existing codebase. This card is the whole discipline: PAGE 2 — the 60-second artboard audit (hunt placeholders · read every button and label off the artboard · re-read the brief's own adjectives) with both real defect autopsies. PAGE 3 — the decision rule: existing-codebase redesign → /design; greenfield → /design + the audit; and pick with the brief's words, not your eyes (the 'calm' dashboard loss, 2–1). PAGE 4 — all 9 round-1 verdicts, the 3–0 redemption flip table, and the fidelity table (97.3/95.5/95.1%, zero missing items). PAGE 5 — the two prompt templates from the test: the /design brief template and the build-from-artboard instruction. PAGE 6 — the honest limits. Every number is a committed receipt from our own run. Independent guide from Hyperautomation Labs — not affiliated with Anthropic.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.