Run on a real page, here is what comes back

A score out of 100, a verdict (generic, mixed or distinctive), and findings ranked by severity, each with the matching code excerpt and a concrete fix.

<style>
  body { font-family: Inter, sans-serif; }
  .btn { border-radius: 9999px; background: linear-gradient(90deg, #7c3aed, #3b82f6); }
</style>
<p class="eyebrow">WELCOME</p>
<h1>Unlock Your Potential</h1>
<a class="btn">Get Started →</a>

generic, 44/100, 4 findings

MEDIUM  Generic default font family (inter)
  fix: Add a second, distinct typeface for headlines and keep the current sans for body/UI text only.

HIGH  Soft tech gradient (purple/blue/pink blend)
  found: linear-gradient(90deg, #7c3aed, #3b82f6)
  fix: Pick one of the two gradient colors as a flat, solid background and drop the other entirely from this element.

HIGH  Eyebrow label directly above a headline
  fix: Delete the label and go straight into the headline.

MEDIUM  AI-cliché vocabulary: unlock
  fix: "unlock" → get / open up

MCP tool check_design, or POST /api/check with url or html.

Cannot do: Reads markup and CSS, not the rendered page. A site that builds its design in script after load will look cleaner here than it is; the vision pass covers that.

The same scored report, limited to wording: cliché vocabulary with replacements, negative-parallel framing, hedge words, heavy em-dash use, problems named without a next step.

We help teams unlock their potential and seamlessly integrate every workflow. It's not about more tools. It's about better outcomes. Our vibrant platform is essentially a game-changer, and there are several errors in the old process that we found missing from most plans.

mixed, 50/100, 4 findings

MEDIUM  AI-cliché vocabulary: vibrant, game-changer, unlock, seamlessly integrate
  fix: "vibrant" → [cut it, name the specific quality instead]; "game-changer" → [cut it and say the specific thing plainly]; "unlock" → get / open up; "seamlessly integrate" → [cut it and say the specific thing plainly]

HIGH  Negative-parallel framing ("It's not X. It's Y.")
  found: It's not about more tools. It's about better outcomes.
  fix: Delete the 'it's not X' half entirely and keep only the direct statement.

LOW  Hedge words: essentially
  fix: Delete "essentially" and reread the sentence, if it still makes sense (it almost always does), leave it out.

HIGH  Diagnosis with no path forward
  fix: For every problem, issue, or error surfaced to a user, pair it with a concrete next step in the same breath, not a separate help page they have to go find.

MCP tool check_copy, or POST /api/check with text.

Cannot do: Matches a fixed word and pattern list. It cannot tell whether the copy suits your brand, and a cliché used on purpose will still be flagged.

Prompt-leak comments such as "as requested" and "I've added", and hedge words in comments, each with the comment text and a rewrite direction.

// As requested, I've added retry logic to the fetch call
async function load(url) {
  // Essentially this just wraps fetch
  for (let i = 0; i < 3; i++) {
    const r = await fetch(url);
    if (r.ok) return r;
  }
}

mixed, 78/100, 2 findings

HIGH  Prompt-leak comments (1)
  found: // As requested, I've added retry logic to the fetch call
  fix: Rewrite each flagged comment to describe the current behavior or the reason for a non-obvious choice, not the act of making the change.

LOW  Hedge words in comments: essentially
  fix: Delete "essentially" or the whole comment if nothing non-obvious is left once they're gone.

MCP tool check_code, or POST /api/check with code.

Cannot do: Reads comments only, in // , /* */ and # styles. A # inside a string can occasionally be mistaken for a comment, and the code itself is never judged.

Extra findings in the same report, written from a Browser Rendering screenshot of your URL that Claude critiques.

No sample is shown for this pass. Every vision run calls Claude on a real key, and we do not run it on the studio's key for a demo, so we have no output of our own to put here.

check_design or POST /api/check with a url, plus either your own Anthropic key in an x-anthropic-key header or a paid Pro key.

Cannot do: Needs a url, not pasted HTML. Free and trialing keys get it only with your own Anthropic key, and Pro on the studio key has a monthly spend allowance.