How to make an AI-built website stop looking AI-built

You can spot a Lovable site, a Bolt site, or a v0-stitched site in about ~3 seconds. Same hero with a centered headline and a soft gradient. Same Stripe-blue feature cards in a 3-column grid. Same testimonial section with avatars from a stock generator. Same "Built with ❤️" footer.

None of this is the tools' fault. The defaults are good defaults. The problem is that everyone ships the defaults.

Below is the pass I run on every AI-built site before it goes live. It's not a redesign — it's a polish loop. Most of these take about 30 minutes each. None of them require rebuilding the app.

The 9 fingerprints, in order of how much they cost you

1. The default font stack (Inter or system + nothing else).
2. The default gradient (cyan-to-purple or blue-to-violet).
3. The default card grid (3 columns, equal weight, tiny icons up top).
4. The default testimonial avatars.
5. The default "as seen in" logo bar.
6. The "all the same spacing" rhythm.
7. Stock illustrations that match nothing.
8. Centered everything.
9. Animations that don't have a reason.

If your site has all 9, you have a Lovable smell. If it has 3 or fewer, it doesn't.

1 — Pick a font that isn't Inter

Inter is fine. Inter is everywhere. That's the problem. Open Google Fonts and pick one of: Geist, Söhne (paid), General Sans, Manrope, Satoshi, Departure Mono (for accents). Pair a sans with a mono and stop there. No serif body fonts unless your brand actually wants editorial.

For Lovable's default Inter setup, switching to Geist is one config line in tailwind.config.ts (or the equivalent in v4). I've never had a project not improve from this single change.

fontFamily: {
  sans: ["Geist", "ui-sans-serif", "system-ui"],
  mono: ["Geist Mono", "ui-monospace"],
}

Add monospace numbers wherever you have prices, dates, or stats. The visual texture difference is huge and it's ~5 minutes of work.

2 — Kill the cyan-to-purple gradient

Gradients aren't the problem. The same two gradients in every AI app are the problem. The cyan-to-violet hero gradient and the pink-to-orange CTA gradient are visually associated with "AI tool default" now.

Fix in order of effort:

• Lowest effort: change one of the gradient stops to an off-brand color (a warm orange, a teal, a deep magenta). One stop different is enough to break the pattern recognition.
• Medium effort: replace the gradient with a single flat brand color and move the visual interest somewhere else (a shader, a photograph, a typography moment).
• High effort: replace the gradient with a real artifact — a screenshot, a real product photo, an animated chart.

The "AI smell" lives in ~80% saturation, ~60% lightness gradients with two adjacent hues. Pick complementary or triadic instead of analogous. Or kill the gradient.

3 — Break the 3-column feature grid

Every AI builder gives you a "Features" section that's a 3-column grid of cards, each with a tiny icon, a <h3>, and 2 lines of text. It's the most copy-pasted layout on the internet right now.

Replacements that take the same time:

• Asymmetric grid — 2 + 1 rows where one card is 2× the size and contains the most important feature.
• Numbered list with prose — drop the cards entirely. Just 01., 02., 03. with a one-paragraph description and a small inline visual. This reads like editorial.
• Conversation flow — "Step 1, then step 2, then step 3" laid out as actual steps with arrows or section breaks. Works great for product onboarding.

The card grid isn't wrong. It's just everywhere. Pick something else ~1 in 3 times.

4 — Replace the avatar testimonials

The default testimonial section pulls avatars from a generator (or worse, leaves the placeholder avatars in). Real humans, real names, real companies, real photos. If you can't get one of those four, write the testimonial as a quote without an avatar — just text and attribution.

Better: link to the source. A tweet screenshot with a working link, a Trustpilot block with a verified badge, a LinkedIn post embed. Anything that says "this is a real person you can verify."

5 — Drop the "as seen in" logo bar (unless it's true)

If you've genuinely been in TechCrunch and Product Hunt and YC, put the logos up. If you haven't, the placeholder logos that come in AI templates make the site feel fake the moment a visitor squints.

Replace with: customer logos (with permission), tool logos you're built with, or nothing. "Nothing" is a real option.

6 — Vary the section rhythm

AI templates produce sections that all have the same vertical padding, the same max-width, the same heading-paragraph-button structure. The page reads as a metronome.

Pick 2–3 sections to break the pattern:

• One section with edge-to-edge content (full bleed image or color).
• One section that's much shorter than the others (a single line of text, large).
• One section with content that floats off-grid (an asymmetric layout).

Variety in vertical rhythm is 80% of why custom sites feel custom.

7 — Cut the stock illustrations

The Undraw / Storyset / Saly figures are doing nothing for your site. They were a 2018 trend and they read as filler now. If you can't replace an illustration with a real screenshot or photograph, delete it.

If you really need a hero visual, options that don't trigger the AI smell:

• A real product screenshot in a frame.
• A WebGL shader (free, MIT-licensed effects abound).
• A gradient mesh tied to your brand colors, not the default.
• A typographic hero — just letters, big.

8 — Stop centering everything

AI templates default to centered headlines, centered paragraphs, centered buttons. It's a safe choice and it reads as "I didn't decide."

Move at least one major section to left-aligned. Hero copy especially benefits from left alignment with a fixed measure (~60ch for body, ~14–22ch for headlines). It feels editorial. It feels human.

9 — Earn every animation

Animations in AI templates are decorative — fade-in on scroll, hover lift on cards, stagger on lists. None of them tell the user anything new.

Audit each animation by asking: does this teach the user something? If the answer is no, remove it. The site will feel faster and more confident.

Animations that earn their keep:

• Reveals on data: a number counts up to a stat. Tells the user "this is the moment you should focus."
• Transitions on state change: a form moves from "filling out" to "submitted." Tells the user the system received their input.
• Cursor-aware effects: hover causes a card to tilt subtly toward the cursor. Adds tactility.

Decorative fade-ins on every section make the site feel slow. Cut them.

The 30-minute checklist

Pull up your AI-built site. Open this list. Spend 30 minutes per item, in order, and stop when you've done 5:

• Swap the font stack.
• Replace or recolor the gradient.
• Pick 1 non-grid layout for your features section.
• Real testimonial sources or remove.
• Logo bar audit (real or remove).
• One section breaks the vertical rhythm.
• Stock illustrations gone.
• Hero left-aligned.
• Every animation has a reason.

After 5 of these, your site will not feel like an AI build. After 9, people will assume you hired a designer.

Prompt pack — when you are still generating with AI

The nine steps above fix a site that already exists. If you are still in the tool (Cursor, Claude, Copilot) and you want the model to ship something that does not read like Slack-default SaaS, you need constraints on the session, not just "make it look good."

Below is a paste-ready scaffold plus workflow rules. The two parts that do the heavy lifting are HARD BANS (closes the easy-path exits) and the SELF-CRITIQUE STEP (forces a second pass). Without both, the model ships its first idea, which is always the safest, dullest one.

1. The anti-Slack prompt template

Fill the bracketed slots. The rest is what fights AI defaults.

ROLE
You are the design director and senior frontend engineer for [brand].
Output target: Awwwards Site of the Day caliber. If a reviewer at Linear,
Vercel, Rauno Freiberg, or Stripe's marketing team would call this generic,
you have failed.

REFERENCE BENCHMARKS (study these before writing a single line)
- linear.app  — typography rhythm, dense feature pages
- vercel.com  — editorial restraint + bold moments
- rauno.me    — asymmetry, monospace details, scroll choreography
- arc.net     — playful but premium, custom illustrations
- framer.com  — interaction polish, type-as-art
- stripe.com/payments — feature page density without clutter

HARD BANS (do not output any of these — they are AI tells)
- Centered hero with [headline] [subhead] [two buttons] [phone mockup]
- 3-column or 6-column "features grid" with icon + title + 2 lines of copy
- Gradient buttons, glow on everything, "shimmer" decoration as content
- Sparkle icons, checkmark glyphs, emoji as iconography (Lucide line icons OK)
- Stock phrases: "empower", "unlock", "next level", "seamless", "elevate",
  "supercharge", "revolutionize", "game-changing"
- Lorem ipsum, fake avatars (UI Faces, Pravatar), fake company logos
- Generic dashboards as filler; placeholder copy
- "Trusted by" logo cloud unless logos are real and earned
- Symmetry where asymmetry would read as edited / curated
- Centered everything — hero, features, testimonials all center-aligned
- The same vertical rhythm in every section (140px padding, 140px padding…)

REQUIRED MOVES (every section must demonstrate at least 2)
- Editorial typography: mixed weights (300/500/700), one display face used
  with intent, monospace for numbers/timestamps/metadata
- Asymmetric layout in at least 50% of sections
- One "bold moment" per fold (a single arresting visual, not five)
- Real content density (a Linear feature page has 600+ words; AI defaults
  to 80 words and 6 icons)
- Numbers, dates, version strings, code excerpts, terminal output where
  natural — proof of craft
- Negative space used aggressively; not everything has a card around it
- At most one rounded-corner radius on the page (e.g. 12px); pick a system

TONE
Voice = [your actual voice — quote 3 sentences from existing copy].
Not corporate. Not breathless. Specific over abstract. Show, don't claim.

DEFINITION OF DONE
- Section reads as if a human design director shipped it
- Could be screenshotted as a portfolio piece
- Removes all stock content; every word/asset is brand-specific
- Mobile gets the same care as desktop, not a stripped-down version
- Passes my self-critique below

SELF-CRITIQUE STEP (you must do this before declaring done)
After implementing, pretend you are Rauno Freiberg reviewing this PR.
List 5 things you would reject and why. Then fix them. Show me the
critique + the fixes in your reply.

CONSTRAINTS
[Paste your brand tokens, fonts, framework, available components]

2. Workflow tactics (matter more than the prompt)

Do not say "build the homepage." That optimizes for complete, not great.

• Section by section, never in bulk. Example: "Design the hero. One section. Show me three radically different directions before writing code."
• Force divergence first. Ask for three distinct concepts in plain text or wireframe before any code. Reject the most obvious one. Pick one of the weirder two.
• Iterate on the bold moment. Example: "The hero's centerpiece — give me five ideas that are not a phone mockup or a dashboard screenshot."
• Design critique loop. After every section: "Pretend you are a senior designer at Apple. What would you reject here?" Then: "Fix all of those."
• Reference images, not adjectives. "Make it look premium" is noise. Drop a screenshot of Linear's homepage and say "this density, this typography rhythm." Cursor and Claude can both ingest images.

3. Model and mode

This matters more than people admit.

• Claude Opus 4.x with high thinking beats Sonnet for raw design taste; the thinking modes spend tokens critiquing their own output, which is what you want here.
• Agent mode beats Ask/Chat for iterative work — the agent can snapshot the running page (browser MCP, Playwright), see what it built, and self-correct. Without sight, it is flying blind.
• GPT-5 high / Codex high is strong for engineering polish (motion, performance) and slightly weaker on pure taste. A practical split: Opus for direction, GPT-class for refactors and perf.
• Avoid fast/cheap models for design work. They are trained to please, which means safe, which means generic.

4. Kill-the-AI-tell weapons

Mid-session prompts that catch the model before it ships Slack-default UI:

• "Show me your typography scale before you write JSX." Forces a system instead of random text-lg / text-xl / text-2xl.
• "What is the one piece of friction in this section that would make a real user pause?" Forces specificity over polish.
• "Rip out 30% of the content. Then 30% more. What survives?" Forces editorial restraint. Models add; they rarely remove.
• "What is the single weirdest thing on this page?" If the answer is "nothing," reject and try again. Strong pages carry one weird, memorable, opinionated element.
• "How does a competitor's mediocre version of this look? Now go in the opposite direction." Negative-space prompting.

5. The reality check

Even with all of the above, AI cannot replace taste. It can execute taste. The prompt's job is to give it taste-by-reference. The sites you name (Linear, Rauno, Vercel, Stripe, and so on) become the anchor for that session. Without them, the model averages every site it has seen — which is generic by definition.

One sentence to staple to every prompt:

Default to the most opinionated, most asymmetric, most editorial choice that still ships in the timebox. If your output could appear on any SaaS website, throw it out and try again.

That line alone moves output from "$5k freelancer safe" toward "$80k senior opinionated." The scaffold above pushes it the rest of the way.

A note on Lovable specifically

Lovable's defaults have improved over time, but the fingerprint is still there in three places:

1. The hero — gradient + centered text + a Stripe-blue button.
2. The pricing section — 3 tiers, middle one highlighted, identical card structure.
3. The footer — small social icons + "Made with Lovable" link (turn this off in settings).

Replace those three sections specifically and your Lovable site will read as custom. {/* DANIEL: drop a screenshot here of a "before/after" Lovable site you've polished. */}

When to hire this out

Polish is the part of the project where most freelancers (and most founders) bail. It's not glamorous, it doesn't ship a feature, and it's hard to measure. But it's what makes the difference between a site that converts and a site that gets bookmarked-but-never-bought.

If you want a Lovable / Bolt / v0 site that stops smelling like AI, tell me about it. I'll usually quote a flat fee for a polish pass and turn it around in ~3–5 days. The whole point of /built-for-you is that you keep the AI's productivity and lose its fingerprint.