The QA gate that caught 47 bugs an AI said didn't exist

KURSYWA PRESS — Sun, 18 Oct 2026 13:07:27 +0000

The QA gate that caught 47 bugs an AI said didn't exist

We let an AI agent build two storefronts. Then, instead of reading its confident summary ("all components render beautifully"), we ran a script that opened the real page in a real browser and asserted things.

It failed 47 times. Every failure was invisible to the model that wrote the code.

Why the model can't see them

Models judge text, not outcomes. A model reviewing its own CSS sees valid syntax and plausible values — it does not see that the buy bar floats off-screen at 360px, that an <img> path 404s behind two working ones, or that the accordion's only animation is prefers-reduced-motion: reduce breaking it for the users who need motion least.

Self-review is the model grading its own homework with the answer key it also wrote.

The ten checks we actually run

Zero horizontal overflow at 360, 768, 1024, 1440, 1920 px.
Every <img> loads (network-level, not alt-text vibes).
Contrast ≥ 4.5:1 on all text against computed backgrounds.
Touch targets ≥ 44px on the primary flow.
Keyboard-reachable focus ring on every interactive element.
Content visible with JS disabled for the core offer.
Reduced-motion page still animates the states that matter (or states nothing).
No emoji bullets, no indigo gradients, no glassmorphism tiles — the slop fingerprint (yes, greppable).
One type scale, one spacing scale — computed styles off the declared tokens only.
Copy passes the specificity test: no sentence that could appear on a competitor's site.

Each is pass/fail. The agent can't argue with an exit code.

The receipts

Buy bar off-screen at 360px: caught 3 times across two builds (different causes, same symptom).
Broken image path: once, on the ONE image that mattered.
Off-scale spacing values: 29 of 47. The model "settled" on no scale at all.
Slop fingerprint hits: 11. All fixed before shipping, none visible to taste-based review.

What changed in our prompting

We stopped asking for polish and started shipping the gate with the brief. The agent sees the ten checks before it writes the first line, and the build fails if it lies. Same model. Different contract.

The full playbook (doctrine + the runnable gate + the teardown of every bug above) and the source bundle are at aiclub1.gumroad.com — and if you've read this far, code PH30 takes 30% off everything until Oct 22.

Free version of the ten-tells list if you want it before money changes hands: aiclub1.gumroad.com/l/tmamtm

We had an AI build two storefronts. Its self-review passed. Its browser QA caught 4 bugs.

KURSYWA PRESS — Thu, 15 Oct 2026 14:00:55 +0000

We had an AI agent build two fictional storefronts one afternoon: a Warsaw watch manufacture and a tiny type foundry. It wrote the pages, ran them in a browser, declared its own work clean. Then we did what most AI-generated work never gets — we graded it.

We call it an acceptance gate: a fixed pass/fail checklist executed against the live page at three breakpoints. Not "make it look premium" — a rubric the model can't charm.

It caught four things the same model had already signed off on:

A hero where a table belonged. The watch page opened with a big emotional statement. What a visitor needed at that moment was a spec table: movement, case size, price. The page looked great and answered nothing.
A buy bar that broke at 375px. The price floated off-screen on mobile. One viewport had been reviewed; the layout was secretly built for exactly that one.
A reduced-motion fallback that wasn't static. prefers-reduced-motion was honored — with a frozen frame that was actually half a rotation mid-spin. A visitor sensitive to motion sees a broken still, not a calm one.
Auto-scaled type that only fit short inputs. The foundry's signage renderer fit its demo word beautifully. Real names overflowed the container. The happy path shipped.

None of this is "AI can't design." All of it is "nobody ran QA on what AI shipped" — and an AI's own review is a rubber stamp, because it grades intentions, not artifacts.

The ten tells

The exercise pushed us to write down what repeats in AI-built UI. Poker logic — tells are the things you do when nobody's watching:

reflex indigo gradients
equal feature tiles, always three
glassmorphism nobody specified
a hero that says nothing at full opacity
copy that's almost-real ("revolutionize your workflow")
the single, centered, impossible-to-miss CTA
emoji bullets wearing a design system's clothes
pricing tables with numbers nobody committed to
states that only exist on the happy path
a layout correct at exactly one viewport

Here is the part that actually matters: the expensive tells are structural, not visual. Indigo is cheap — one line of prompt kills it before code exists, which is also why classifiers trained on visual tells chase a moving target. "Shipping a hero on a screen that needs a table" survives every color change. That is the class worth putting on a checklist.

What we did with it

Two artifacts, boring on purpose:

a ~1KB pre-prompt of constraints that kills the visual tells at generation time
a pass/fail browser gate at 3 breakpoints whose failure screenshots get saved as receipts — the teardowns in our playbook are exactly this, bugs included, because the bugs are the product

We publish as KURSYWA PRESS. Disclosure, since this is a shop: the full doctrine is a $29 PDF; the 2-page extract is free → https://aiclub1.gumroad.com/l/tmamtm

Curious what the gate flagged on your AI-built page? Drop a screenshot description in the comments — worst structural tell wins a review.

DEV Community: KURSYWA PRESS

The QA gate that caught 47 bugs an AI said didn't exist

The QA gate that caught 47 bugs an AI said didn't exist

Why the model can't see them

The ten checks we actually run

The receipts

What changed in our prompting

We had an AI build two storefronts. Its self-review passed. Its browser QA caught 4 bugs.

The ten tells

What we did with it