Prompt Patterns That Work
Concrete templates and anti-patterns from real Lovable sessions. The phrasing that gets results, and the phrasing that wastes a prompt.
You will write 30-60 prompts in a typical Project 1 session. Some of them will work in one shot. Some will need three tries. The difference is rarely the AI's fault — it is the phrasing.
This page is a reference. Skim it now. Come back to it when a prompt produces something disappointing.
Pattern 1 — The Specific Single Change
Good.
"Make the hero background indigo-900 (#312E81) and the headline text white."
Bad.
"Make the hero look better."
The good version gives the AI a specific element (hero background, headline text), a specific value (indigo-900, white), and a clear scope (just those two things). The AI can do exactly that and nothing else.
The bad version asks for "better" — the AI has to guess what you mean. Often it will redesign the entire hero, breaking things you liked.
Rule: every prompt should answer three questions — what element, what change, to what value.
Pattern 2 — The Reference Comparison
Good.
"Make the spacing in the branches section more like notion.so's homepage — generous padding around each card, 24px gap between cards."
The AI knows what notion.so looks like. Giving a reference cuts through ambiguity. Combine with a specific measurement (24px gap) for precision.
Bad.
"More spacing."
The AI does not know how much "more" means. It might add 4px or 80px.
Pattern 3 — The Multi-Step Build
When you need a feature with multiple parts, don't ask for all of them at once. Sequence them.
Bad — one giant prompt.
"Add a 4-step admissions process with icons, then a contact form with validation and a success message, then a section showing tuition fees, then a parent testimonial carousel."
The AI will do all four — and at least one will be subtly broken or look unrelated to the rest.
Good — sequenced prompts.
Prompt 1: "Add a section called 'How to apply' with 4 numbered steps. Each step has an icon, a heading, and one sentence of description."
(see result)
Prompt 2: "Below the steps, add a contact form — name, parent email, child's grade, and message. Show a green success message on submit. Form doesn't need to actually send anything."
(see result)
Prompt 3: "Replace the tuition fee placeholder with a small section that says 'Tuition fees are shared on inquiry. Please fill the form above.'"
You get something coherent. Each step builds on a stable foundation.
Pattern 4 — The Anchor Word
Lovable tends to lose track of which element you mean when there are similar things on the page. Use anchor words.
Vague.
"Make the title bigger."
The AI might enlarge the page title, the hero title, the section title — pick one randomly.
Anchored.
"Make the hero title bigger — about 60px on desktop, 36px on mobile."
Now the AI knows exactly which title and roughly how big.
Other useful anchors:
- "the indigo button in the hero" (not just "the button")
- "the second section, with the 3 cards" (not just "the section")
- "the photo of children in the lab" (not just "the photo")
Pattern 5 — The Negative Constraint
Sometimes telling the AI what not to do is the fastest way to get what you want.
"Add a testimonials section with 3 short parent quotes. Do not use a carousel. Make it a simple 3-column grid that wraps to 1 column on mobile."
The AI's default for testimonials is often a carousel. Saying "not a carousel" upfront saves an iteration.
Other negative constraints:
- "Do not change anything outside the hero."
- "Keep the existing colour palette — don't introduce any new colours."
- "Don't add any animation."
- "Keep the existing image — only change the text."
Pattern 6 — The Verbose Tone Direction
When you want the AI to write copy (the actual sentences on the page), pay for it with words.
Thin.
"Write about the school."
Result: generic mission statement that could be any school anywhere.
Rich.
"Write 3 short paragraphs about Aurora Public School. Tone: warm but not saccharine. Audience: a parent considering schools for their 7-year-old. Mention robotics from Grade 4 (without overselling it), the family feel because all 3 campuses are within Mumbai, and that teachers stay for years. Don't use the phrases 'holistic development' or 'world-class' — they're cliches."
Result: copy that sounds like an actual school wrote it.
The pattern is audience + tone + must-mention + must-avoid.
Pattern 7 — The "Why" Hint
Sometimes the AI does something that looks wrong because it does not know your goal. Telling it why you want something often produces a different and better result.
Without why.
"Make the contact section more prominent."
Result: bigger headings, more padding, maybe a coloured background.
With why.
"Make the contact section more prominent — most visitors to a school site are parents trying to figure out who to talk to. The phone number should be the easiest thing to find on the whole page."
Result: a sticky bottom bar with the phone number, or a prominently coloured CTA box. The AI now knows prominence means findability, not just bigness.
Pattern 8 — The Lock-In
Once you get something right, say so, and ask the AI to lock it in before you change anything else.
"The current hero is good — keep it exactly as-is. Now make changes to the about section below the hero."
This dramatically reduces the chance the AI "improves" something you didn't want changed. Lovable doesn't have a literal lock feature, but mentioning what to preserve carries weight.
Anti-Patterns to Avoid
1. Re-explaining the entire site every prompt
Lovable remembers context within a session. You do not need to start every prompt with "we're building a school landing page for Aurora Public School..." — it knows.
2. Asking for huge changes
"Redesign the entire site to look more modern."
This usually replaces things you liked. If most of the site is close to right, change one section at a time.
3. Using vague qualitative words
| Vague | Specific |
|---|---|
| "more vibrant" | "the indigo more saturated, like #4F46E5" |
| "cleaner" | "remove the borders on the cards; use shadow only" |
| "more professional" | "use a serif font for headings only (e.g. Cormorant Garamond)" |
| "feels off" | "the hero text is hard to read against the photo — add a dark overlay" |
4. Multiple unrelated changes in one prompt
"Fix the spacing in the hero, change the about copy, add a new section, and make the footer dark."
Even if each one is specific, doing all four at once is risky. The AI will get some right and some wrong, and you won't easily know which.
5. Asking for things outside vibe-coding's strengths
"Make the contact form actually send an email to admissions@aurora-school.com."
Lovable can scaffold this, but the wiring (SMTP, API keys, deliverability) is fragile. Save it for Project 2 where you do it properly with Resend.
A Working Session, Compressed
Here is the actual sequence of prompts from a real successful Project 1 session, abbreviated:
- "A landing page for Aurora Public School..." (the big setup prompt)
- "The hero photo is generic. Use a brighter classroom shot with children."
- "Move the 'Plan a Visit' button below the tagline, not next to it."
- "Make the button amber (#F59E0B) with dark text."
- "The headline 'Where curious children grow' is too small. Increase to about 56px on desktop."
- "Below the hero, add the 3-paragraph about. Tone: warm, parent-direct, mention robotics."
- "The about section is too narrow. Make it 720px max-width, centred."
- "Now add 'Our Branches' below — 3 cards in a row: Andheri, Powai, Thane."
- "Each branch card should have a small map pin icon, address, phone."
- "The branch cards look flat. Add a soft shadow on hover and a subtle border."
- "Footer is missing. Add: phone, email, address, and a single-line copyright."
- "Footer is too dark — match the hero's indigo, not pure black."
- "Check mobile. Fix anything that wraps badly."
- "The 'Plan a Visit' button looks too small on mobile. Make it full-width below the headline on mobile only."
- "Looking good. Add a second page at /branches with full cards per branch."
- ... (15 more iterations) ...
- "Publish."
Notice: no prompt is more than two sentences. Every prompt addresses one thing. The work happens in dozens of small steps, not a few big ones.
This is the rhythm. You will be in this rhythm for 4-6 hours total. Then you publish.