A post went around X this week that just said every website UI is vibe coded now. Nobody in the replies argued. Around the same time, a demo got passed around of an MCP server that feeds a coding agent curated designs from real websites, so the agent has something better to copy than its own defaults. People shared it because they already have the problem it fixes. You open a new SaaS landing page and you can tell what made it before the page finishes loading.
I build with these tools every day, and I don't think the issue is that AI-generated sites are ugly. Most of them look fine. The problem is that they all look the same kind of fine, and visitors notice. A page that looks like a template reads like a weekend project, even when the product behind it is good.
"It looks AI-made" is a vibe, and you can't review a vibe in a pull request. So I tried to turn it into something you can check: a small linter that looks for the default look in a project's HTML, JSX, CSS and Tailwind classes, prints what it found, and gives a score.
The fingerprints
These are my observations from looking at a lot of generated marketing pages, plus my own first drafts. I haven't measured any of this. It's a pattern list, and you'll recognize most of it:
-
One font, and it's Inter. Or
system-ui, or Tailwind'sfont-sanswith nothing set. No display face, no contrast between headings and body text. -
A purple-to-blue gradient hero.
from-purple-600 via-indigo-600 to-blue-500, or the same colors written as a CSSlinear-gradient. -
Glassmorphism.
backdrop-bluronbg-white/10, usually on the sticky nav and a "Now with AI" pill. -
Three rounded cards in a row.
md:grid-cols-3,rounded-2xl, a soft border, a hover lift. - Emoji or lucide icons as the card art. ⚡ for speed, 🔒 for security, 📈 for insights.
-
A grey "Trusted by" logo bar.
grayscale opacity-60, often showing logos the company has no relationship with. - A centered hero with two buttons. "Get started" and "Learn more", side by side.
- Launch copy that could describe any product. Supercharge, unlock, seamless, effortless, empower, elevate, all-in-one, streamline.
-
The same spacing everywhere. Every section gets
py-24, so the page has no rhythm. It's just one block after another. -
Testimonials that look made up. "Sarah Johnson, CEO at Acme", a
pravatar.ccavatar, five star glyphs.
None of these is wrong on its own. Inter is a great typeface, and three cards are sometimes the right layout. The fingerprint is all of them showing up together, because together they mean nobody made a design decision.
The fixture
I didn't scan anyone's real site. I wrote a fixture for a made-up invoicing tool called Ledgerly, the way a coding agent tends to write it when you prompt "build a modern landing page for an invoicing SaaS" and give it nothing else: an index.html with Tailwind classes, a small styles.css, and a src/Testimonials.jsx component. I put the fingerprints in on purpose, so the "before" score below is expected to be high. The useful part is what happens after.
sameness-lint
It's one Python file, stdlib only, about 140 lines. Each rule has a weight, and the weights add up to 100. A rule adds its weight once if it finds anything. Buzzwords cost 3 points each, up to the rule's 12-point cap. Here's the core:
RULES = { # id: (weight, label)
"default-font": (12, "Only Inter/system sans in the whole project"),
"purple-gradient":(14, "Purple-to-blue gradient"),
"glass": (8, "Glassmorphism (backdrop blur on translucent white)"),
"card-grid-3": (10, "Three-up rounded card grid"),
"icon-cards": (8, "Emoji / lucide icons as feature-card art"),
"logo-bar": (8, "Greyed-out 'Trusted by' logo bar"),
"centered-hero": (10, "Centered hero with two CTAs"),
"buzzwords": (12, "Generic launch copy"),
"same-rhythm": (8, "Every section uses the same vertical padding"),
"stock-proof": (10, "Placeholder-looking testimonials"),
}
# gradients: Tailwind stops, or CSS hex stops that all sit in the violet/blue band
for m in re.finditer(r"from-(purple|violet|indigo|fuchsia)-\d+[^\"']*?to-(blue|indigo|cyan|sky|purple|pink)-\d+", low):
add("purple-gradient", m.start(), m.group(0))
for m in re.finditer(r"linear-gradient\(([^)]*)\)", low):
hs = [hue(h) for h in re.findall(r"#[0-9a-f]{6}\b|#[0-9a-f]{3}\b", m.group(1))]
if len(hs) >= 2 and all(200 <= h <= 300 and s > 0.4 for h, s in hs):
add("purple-gradient", m.start(), m.group(0))
# centered hero: text-center near the <h1>, and at least two generic CTAs after it
for m in re.finditer(r"<h1\b", low):
before, after = low[max(0, m.start() - 500): m.start()], low[m.start(): m.start() + 1500]
if "text-center" in before + after[:200] and len(re.findall(CTA, visible(after))) >= 2:
add("centered-hero", m.start(), "text-center h1 + ...")
# rhythm: if 75%+ of <section>s share one py-* value, the page has no rhythm
if sections >= 3 and len(where) / sections >= 0.75:
hits["same-rhythm"].append(f"{len(where)}/{sections} sections use {py}")
The font rule collects every font-family declaration and every Google Fonts family= in the project. If all of them are in the default sans set (Inter, system-ui, Roboto, Helvetica, Arial and friends), it flags the project. The buzzword rule strips tags and class attributes first, so class="seamless-grid" doesn't count and visible copy does. A file can opt out of a rule with a sameness-lint-disable <rule> comment. That's on purpose: the opt-out shows up in the diff, where a reviewer can push back on it.
Full sameness_lint.py (about 140 lines, stdlib only)
#!/usr/bin/env python3
"""sameness-lint: flag the "AI default" look in a marketing site's HTML/JSX/CSS.
Heuristics, not taste. Each rule is a fingerprint you see a lot in generated sites."""
import re, sys, json, colorsys
from pathlib import Path
EXTS = {".html", ".htm", ".jsx", ".tsx", ".js", ".ts", ".vue", ".svelte", ".css", ".scss"}
SKIP = {"node_modules", ".git", "dist", "build", ".next"}
DEFAULT_FONTS = {"inter", "system-ui", "-apple-system", "blinkmacsystemfont", "sans-serif",
"ui-sans-serif", "segoe ui", "roboto", "helvetica", "arial", "helvetica neue"}
BUZZ = ["supercharge", "unlock", "seamless", "effortless", "revolutioni", "empower", "elevate",
"next-level", "game-chang", "cutting-edge", "all-in-one", "streamline", "unleash",
"transform your", "10x", "at scale", "like never before"]
CTA = r"(get started|start (?:for )?free|try (?:it )?free|book a demo|learn more|see how|watch demo)"
STOCK = r"(i\.pravatar\.cc|randomuser\.me|ui-avatars\.com|\bjohn doe\b|\bjane doe\b|\bacme\b|lorem ipsum|★★★★★)"
EMOJI = re.compile("[\U0001F300-\U0001FAFF\u2600-\u27BF]")
RULES = { # id: (weight, label)
"default-font": (12, "Only Inter/system sans in the whole project"),
"purple-gradient":(14, "Purple-to-blue gradient"),
"glass": (8, "Glassmorphism (backdrop blur on translucent white)"),
"card-grid-3": (10, "Three-up rounded card grid"),
"icon-cards": (8, "Emoji / lucide icons as feature-card art"),
"logo-bar": (8, "Greyed-out 'Trusted by' logo bar"),
"centered-hero": (10, "Centered hero with two CTAs"),
"buzzwords": (12, "Generic launch copy"),
"same-rhythm": (8, "Every section uses the same vertical padding"),
"stock-proof": (10, "Placeholder-looking testimonials"),
}
def files(root):
for p in sorted(Path(root).rglob("*")):
if p.is_file() and p.suffix in EXTS and not SKIP & set(p.parts):
yield p
def line_of(text, idx):
return text.count("\n", 0, idx) + 1
def visible(text): # strip markup but keep newlines so line numbers still work
keep = lambda m: "\n" * m.group(0).count("\n") + " "
t = re.sub(r"<(script|style)\b.*?</\1>", keep, text, flags=re.S | re.I)
t = re.sub(r"\{/\*.*?\*/\}|className=\"[^\"]*\"|class=\"[^\"]*\"", keep, t, flags=re.S)
return re.sub(r"<[^>]+>", keep, t)
def hue(hexstr):
h = hexstr.lstrip("#")
if len(h) == 3: h = "".join(c * 2 for c in h)
r, g, b = (int(h[i:i + 2], 16) / 255 for i in (0, 2, 4))
hh, l, s = colorsys.rgb_to_hls(r, g, b)
return hh * 360, s
def scan(root):
hits = {k: [] for k in RULES}
fonts, rhythm = set(), {}
for p in files(root):
rel, text = str(p.relative_to(root)), p.read_text(errors="ignore")
low = text.lower()
off = set(re.findall(r"sameness-lint-disable ([\w-]+)", text)) # per-file opt-out, reviewed in PR
def add(rule, idx, ev, src=text):
if rule in off: return
item = f"{rel}:{line_of(src, idx)} {ev.strip()[:70]}"
if item not in hits[rule]: hits[rule].append(item)
# fonts: CSS font-family, Google Fonts links, Tailwind font config
for m in re.finditer(r"font-family\s*:\s*([^;}{]+)", low):
fonts |= {f.strip(" '\"") for f in m.group(1).split(",") if f.strip(" '\"") and "var(" not in f}
for m in re.finditer(r"fonts\.googleapis\.com/css2?\?([^\"']+)", low):
fonts |= {f.split(":")[0].replace("+", " ") for f in re.findall(r"family=([^&\"']+)", m.group(1))}
# gradients: Tailwind stops or CSS hex stops in the violet/blue band
for m in re.finditer(r"from-(purple|violet|indigo|fuchsia)-\d+[^\"']*?to-(blue|indigo|cyan|sky|purple|pink)-\d+", low):
add("purple-gradient", m.start(), m.group(0))
for m in re.finditer(r"linear-gradient\(([^)]*)\)", low):
hs = [hue(h) for h in re.findall(r"#[0-9a-f]{6}\b|#[0-9a-f]{3}\b", m.group(1))]
if len(hs) >= 2 and all(200 <= h <= 300 and s > 0.4 for h, s in hs):
add("purple-gradient", m.start(), m.group(0))
for m in re.finditer(r"backdrop-blur[\w-]*|backdrop-filter\s*:\s*blur", low):
ctx = low[max(0, m.start() - 200): m.end() + 200]
if re.search(r"bg-white/\d+|rgba\(255,\s*255,\s*255,\s*0?\.\d+\)", ctx):
add("glass", m.start(), m.group(0))
for m in re.finditer(r"(?:md:|lg:)?grid-cols-3", low):
if low[m.end(): m.end() + 1500].count("rounded-2xl") + low[m.end(): m.end() + 1500].count("rounded-3xl") >= 3:
add("card-grid-3", m.start(), m.group(0) + " + rounded-2xl cards")
for m in re.finditer(r"from ['\"]lucide-react['\"]|data-lucide=", low):
add("icon-cards", m.start(), m.group(0))
for m in re.finditer(r"<(?:h3|div|span)[^>]*>\s*(" + EMOJI.pattern + ")", text):
add("icon-cards", m.start(), "emoji icon " + m.group(1))
for m in re.finditer(r"(trusted by|loved by|used by teams at|backed by)", low):
if re.search(r"grayscale|opacity-(?:[1-6]0)|text-(?:gray|slate|zinc)-[34]00", low[m.start(): m.start() + 1200]):
add("logo-bar", m.start(), m.group(0) + " + greyed logos")
for m in re.finditer(r"<h1\b", low):
before, after = low[max(0, m.start() - 500): m.start()], low[m.start(): m.start() + 1500]
if "text-center" in before + after[:200] and len(re.findall(CTA, visible(after))) >= 2:
add("centered-hero", m.start(), "text-center h1 + " + ", ".join(sorted(set(re.findall(CTA, visible(after))))))
if p.suffix != ".css":
vis = visible(text).lower()
for w in BUZZ:
for m in re.finditer(re.escape(w), vis):
add("buzzwords", m.start(), f'"{w}"', vis)
for m in re.finditer(r"<section\b[^>]*class(?:name)?=\"([^\"]*)\"", text, re.I):
pys = re.findall(r"\bpy-\d+\b", m.group(1))
if pys: rhythm.setdefault(pys[-1], []).append(f"{rel}:{line_of(text, m.start())}")
for m in re.finditer(STOCK, low):
add("stock-proof", m.start(), m.group(0))
if not fonts or fonts <= DEFAULT_FONTS:
hits["default-font"].append("font stack: " + (", ".join(sorted(fonts)) or "none declared (Tailwind font-sans)"))
sections = sum(len(v) for v in rhythm.values())
for py, where in rhythm.items():
if sections >= 3 and len(where) / sections >= 0.75:
hits["same-rhythm"].append(f"{len(where)}/{sections} sections use {py}")
return hits
def score(hits):
s = 0
for k, (w, _) in RULES.items():
n = len(hits[k])
if k == "buzzwords": s += min(w, 3 * n) # 3 points per phrase, capped
elif n: s += w
return s
def main():
args = [a for a in sys.argv[1:] if not a.startswith("--")]
root = Path(args[0] if args else ".")
limit = int(next((a.split("=")[1] for a in sys.argv if a.startswith("--max=")), 101))
hits = scan(root); total = score(hits)
if "--json" in sys.argv:
print(json.dumps({"score": total, "hits": hits}, indent=2)); return
print(f"sameness-lint {root}\n")
for k, (w, label) in RULES.items():
if hits[k]:
print(f"✗ {label} [{k}, {len(hits[k])}]")
for ev in hits[k][:3]: print(f" {ev}")
if len(hits[k]) > 3: print(f" … +{len(hits[k]) - 3} more")
clean = [k for k in RULES if not hits[k]]
if clean: print(f"✓ clean: {', '.join(clean)}")
print(f"\nsameness score: {total}/100 (lower is more yours)")
if total >= limit:
print(f"FAIL: score {total} >= --max={limit}"); sys.exit(1)
if __name__ == "__main__":
main()
Run 1: the default look
Here's the real output on the fixture, run with --max=30 the way it would run in CI:
$ python3 sameness_lint.py fixtures/before --max=30
sameness-lint fixtures/before
✗ Only Inter/system sans in the whole project [default-font, 1]
font stack: -apple-system, inter, sans-serif, system-ui
✗ Purple-to-blue gradient [purple-gradient, 2]
index.html:19 from-purple-600 via-indigo-600 to-blue-500
styles.css:2 linear-gradient(135deg, #7c3aed 0%, #4f46e5 50%, #3b82f6 100%)
✗ Glassmorphism (backdrop blur on translucent white) [glass, 3]
index.html:12 backdrop-blur-md
index.html:21 backdrop-blur
styles.css:3 backdrop-filter: blur
✗ Three-up rounded card grid [card-grid-3, 1]
index.html:42 md:grid-cols-3 + rounded-2xl cards
✗ Emoji / lucide icons as feature-card art [icon-cards, 7]
index.html:14 emoji icon ✨
index.html:21 emoji icon 🚀
index.html:44 emoji icon ⚡
… +4 more
✗ Greyed-out 'Trusted by' logo bar [logo-bar, 1]
index.html:32 trusted by + greyed logos
✗ Centered hero with two CTAs [centered-hero, 1]
index.html:22 text-center h1 + get started, learn more
✗ Generic launch copy [buzzwords, 17]
index.html:6 "supercharge"
index.html:22 "supercharge"
index.html:23 "unlock"
… +14 more
✗ Every section uses the same vertical padding [same-rhythm, 1]
5/5 sections use py-24
✗ Placeholder-looking testimonials [stock-proof, 6]
index.html:34 acme
src/Testimonials.jsx:4 acme
src/Testimonials.jsx:4 i.pravatar.cc
… +3 more
sameness score: 100/100 (lower is more yours)
FAIL: score 100 >= --max=30
[exit 1]
All ten rules fired, so the fixture scored 100 out of 100. Again, that's by design. I wrote the fixture to hit every rule. The evidence lines are the useful part: 17 buzzword hits (five of them in the testimonial quotes), three placeholder avatars, and all five sections using py-24.
The scanner also over-counts in one place. The ★★★★★ string matched twice: once as "stock proof" and once as an "emoji icon", because U+2605 falls in the symbol range I scan for emoji. The score doesn't change, because each rule only adds its weight once, but don't treat the hit counts as precise measurements.
The fix: write the brief before the first prompt
What actually changed the output wasn't a better prompt. It was a file the agent reads before it writes any markup. I keep a DESIGN.md in the repo root and point the agent at it in the first message ("Read DESIGN.md. Do not use anything it bans."). Here's the template, filled in for Ledgerly:
# DESIGN.md — Ledgerly marketing site
## Who it's for
Freelancers and two-to-five person studios who invoice 5–40 clients a month
and hate chasing late payments. Not "modern teams". Not enterprises.
## Type
- Display: Fraunces 600, optical size on, tight tracking (-0.02em) at 48px+
- Body/UI: IBM Plex Sans 400/500, 17px base, 1.55 line height
- Numbers: IBM Plex Mono for amounts, dates, invoice IDs
- Never: Inter, system-ui as the only stack
## Palette (no gradients on backgrounds)
| Role | Hex |
|-------------|---------|
| Paper (bg) | #F6F1E7 |
| Ink (text) | #1B1B18 |
| Ledger green (primary) | #1F5C4A |
| Stamp red (overdue, one accent per screen) | #C8452B |
| Rule lines | #D8CFBE |
## Spacing and rhythm
8px base. Sections do not share one padding: hero 112/64, proof 48, features 96,
closing 64. Content column 680px for prose, 1040px for layouts.
Corners: 4px on inputs and buttons, 0 on the invoice. No rounded-2xl cards.
## Signature move
The hero *is* an invoice: a ruled, left-aligned paper document with a real
line-item table and a red "OVERDUE 14 DAYS" stamp, next to the headline.
Every other section borrows the ruled-line motif. No floating blobs, no glass.
## Components
- One primary CTA per screen. Secondary actions are text links.
- Features as a numbered list with a one-line outcome each, not icon cards.
- No emoji. Icons only when they label a control.
## Voice
- Say what happens: "Ledgerly emails the reminder on day 3, 7 and 14."
- Use the customer's nouns: invoice, client, retainer, late fee.
- Numbers only if we can source them. No invented stats.
- Banned: supercharge, unlock, seamless, effortless, empower, elevate,
revolutionize, game-changer, cutting-edge, all-in-one, streamline,
next-level, "for modern teams", "like never before".
## Social proof
Real customer, real name, permission on file, or no testimonial section.
No placeholder avatars, no five-star glyphs, no fake logo bars.
## References (look at these, don't copy them)
- A paper invoice from an actual accountant
- Stripe's docs for table density
- A newspaper classifieds page for ruled-line rhythm
The section that does the most work is the signature move. Fonts and hex codes stop the page from looking generic, but one specific layout idea tied to the product (here, "the hero is an overdue invoice") is what makes the agent stop reaching for a centered hero. If you can't name one, look at real sites in your category before you prompt. A list of real business website examples is a better starting point than "make it modern". If you're using a references MCP server like the one in that demo, give it two or three sites you actually like and tell the agent what to take from each one: the density, the type contrast, or the way the proof section is laid out. Don't ask it to copy the whole page.
The other section that earns its place is Social proof. The before fixture had invented testimonials because the prompt asked for a "complete" page. If you don't have a real quote yet, leave the section out. A page with no testimonials looks more honest than one with Sarah Johnson from Acme.
Run 2: after the brief
I rebuilt the page against the brief. The hero is now a left-aligned Fraunces headline next to a ruled invoice with the stamp. The logo bar became a plain mono strip of facts (reminders on day 3/7/14, card or SEPA, CSV export). Features became a numbered list, and section padding varies. Then I added a src/Pricing.jsx component the way it usually arrives: generated in a separate session that never saw the brief.
$ python3 sameness_lint.py fixtures/after --max=30
sameness-lint fixtures/after
✗ Emoji / lucide icons as feature-card art [icon-cards, 1]
src/Pricing.jsx:1 from "lucide-react"
✗ Generic launch copy [buzzwords, 1]
src/Pricing.jsx:7 "seamless"
✓ clean: default-font, purple-gradient, glass, card-grid-3, logo-bar, centered-hero, same-rhythm, stock-proof
sameness score: 11/100 (lower is more yours)
[exit 0]
The score went from 100 to 11, and both remaining hits are in the component that never saw DESIGN.md. That's the most useful thing this run showed. The brief works on the session that reads it, and new sessions start from defaults again.
The two hits aren't equally important:
-
"seamless"in a pricing bullet is a real regression. I rewrote it as "Nightly CSV push to your accountant's inbox", which says what the feature does. - The
lucide-reactimport is arguable. ThoseCheckicons mark list items in a pricing table, which the brief allows ("icons only when they label a control"). So the file gets asameness-lint-disable icon-cardscomment with the reason next to it, and a reviewer can disagree in the PR.
After both changes:
$ python3 sameness_lint.py fixtures/after-review --max=30
sameness-lint fixtures/after-review
✓ clean: default-font, purple-gradient, glass, card-grid-3, icon-cards, logo-bar, centered-hero, buzzwords, same-rhythm, stock-proof
sameness score: 0/100 (lower is more yours)
[exit 0]
A 0 doesn't mean the page is good. It means the page avoids ten specific defaults. Taste is still your job, and the linter can't tell you whether the invoice hero actually works for a freelancer. It just stops the easy defaults from slipping back in.
Make it a pre-merge check
Linters only help if they run without anyone remembering to run them. Mine runs on every PR that touches the marketing site:
# .github/workflows/sameness.yml
name: sameness-lint
on:
pull_request:
paths: ["site/**"]
jobs:
lint:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: actions/setup-python@v5
with: { python-version: "3.12" }
- run: python3 tools/sameness_lint.py site --max=30
A few things I'd tell anyone copying this:
-
Set the threshold from your own baseline. Run it on your current site first. If you score 40, set
--maxto 45 and tighten it over time. 30 is just what I picked for this fixture. -
Tune the rules to your brief. If your brand really does use Inter, delete that rule. The rules should enforce your
DESIGN.md, not my taste. -
Add your agent's tics. Every model has its own favorite phrases. When you catch a new one in review, add it to
BUZZand the brief's banned list in the same commit. - Keep the opt-out visible. A per-file disable comment shows up in the diff. A global config flag doesn't.
Where this fits
It doesn't matter much whether the site comes from Cursor, Claude Code or an AI website builder. The first draft will lean toward the average of everything the model has seen. That's what "default" means. That's true of every tool, PolyDraft included, which is why the brief matters more than the prompt. If you're comparing tools, generate a page from the same brief in each one and run a check like this on the output. It's a useful extra row on a list of what to check when evaluating an AI website builder. And if your "site" is really one page right now, decide whether it should stay a landing page or grow into a website (the difference matters for what you build next) before you write the brief.
What's the tell you spot first on an AI-built site? I'll add good ones to the rule list.



Top comments (0)