Writing effective prompts

How-to guide

Get a site that fits your business by giving the AI three things: what you sell, who buys it, and the specific structure you want. MaShop reads your prompt, builds a complete Next.js site, and commits it to your repo. Vague prompts produce generic layouts; concrete prompts produce the right pages, the right vocabulary, and the right data model. This page shows you how to write the prompt and how to iterate after the first build.

Describe your business concretely

State what you sell in the first sentence. MaShop has no fixed list of business types. The AI treats common shapes (a shop, a restaurant with reservations, a service with a booking calendar, an event with tickets, a subscription) as inspiration, not a menu you pick from. When your business sits outside those shapes, the AI infers the right pages, sections, vocabulary, and database tables for what you actually offer instead of forcing an e-commerce template onto it.

Because the AI infers structure from your words, the words carry weight. Name the things you sell with the real noun: dishes, sessions, tickets, plans, courses, not "products" unless you literally sell products. The generation engine uses that vocabulary to pick which sections to build and what to call your catalog.

A thin prompt and a concrete prompt produce different sites:

Thin prompt	Concrete prompt
Make me a coffee shop site.	An online roastery selling single-origin coffee beans. Customers filter by roast level and origin, pick a grind, and buy. Each bag has photos, a tasting note, and a price per 250g. Add a subscription option for monthly delivery and an email signup at the bottom.
Build a site for my gym.	A photography studio that sells 1-hour booking sessions. Clients see available time slots for the week, pick a date and time, and pay. Show a portfolio gallery on the home page and a team page with each photographer's specialty.

The concrete prompts name the items (bags of coffee, booking sessions), the buyer action (filter and buy, pick a slot and pay), and the supporting sections (subscription, email signup, portfolio, team). The AI builds against those specifics.

Give the AI the specifics it can act on

A strong first prompt covers five things. Each one maps to a real decision the generation engine makes.

• What you sell. The item and its real noun. This sets the catalog vocabulary and which section patterns the AI reaches for.
• Who buys it. Your audience shapes tone and layout density.
• The pages you need. Name them if you have an opinion (Home, Shop, About, Contact, a booking page, a menu page). If you do not, the AI always builds a Home page and adds the pages that fit your business.
• Key sections. The site is assembled from named section patterns. The shared library includes a header, hero, features, testimonials, a call-to-action, stats, newsletter, footer, FAQ, gallery, about, contact, and pricing block. Business-fit sections cover product grids and a slide-out cart, service lists and availability calendars, menu sections grouped by category, ticket tiers, pricing tiers with a feature comparison, and download sections. Ask for the ones you want by name.
• Item details. Tell the AI what each item carries: photos, a description, price, variants like size or color, a duration for a service, dietary badges for a dish.

Two facts shape what you should and should not put in a prompt:

• You do not write currency or prices into the design. Every data-displaying section reads your catalog and your store settings at runtime, and prices render in the currency from your settings. State the currency once in your project setup or your prompt; never hardcode a symbol into a section description.
• Pick your payment provider explicitly. The AI wires the checkout to the provider you name. Say "use Stripe for payments" or name another provider. Do not assume a default.

Let the first prompt build the whole site

For the initial build, hand the AI one rich prompt and let it generate everything at once. The engine has a combined path that produces the full site and its SEO (meta tags, JSON-LD, sitemap) in parallel, which is the preferred route for a first build and runs 2 to 3 times faster than generating the site and then optimizing SEO in separate steps.

A complete first prompt for a clothing brand:

An online store for a sustainable fashion brand. Show products grouped by
category: shirts, trousers, accessories. Each product has photos, a description,
sizes, and color options. Add a persistent cart with a slide-out drawer, a
discount-code field at checkout, and a newsletter signup in the footer. Use
Stripe for payments. Tone: minimal and premium.

From that one message the AI builds the home hero, a category-filtered product grid, the cart drawer, the checkout flow, the Supabase tables behind them, the coupon logic, and the payment webhook. It commits the deployable files to your repo as it goes.

A complete first prompt for a restaurant:

A neighborhood Italian restaurant with online reservations. The menu is grouped
into sections: antipasti, pasta, mains, desserts. Each dish shows a name, a short
description, a price, and dietary badges (vegetarian, gluten-free). Add a
reservation form with date, time, and party size on a dedicated page. Warm, dark
theme. Show opening hours and the address.

Note the vocabulary: "dishes" and "menu sections," not "products." That is what tells the AI to build a menu grouped by category rather than a shopping grid.

Iterate one section at a time

After the first build, make follow-up changes small and targeted. Changing one section is faster and safer than rebuilding the whole site, because the AI can edit a single section in place without regenerating everything. Name the section you mean.

Section-level edits, one per prompt:

Change the header background to dark navy and make the logo bigger.

Rewrite the hero headline to lead with free shipping over 50 EUR.

Add an FAQ section before the footer with five questions about returns and delivery.

Each of these touches one section. The AI keeps the rest of the page visually consistent and commits only the changed file.

Some follow-ups are not design edits at all; they are catalog and configuration actions. When you ask the AI to add a product, create a category, set up a coupon, write a blog post, or change a store setting, it runs that as a dashboard action against your data. Real examples it can act on:

• Catalog: "Add a product called Ethiopia Yirgacheffe at 18 EUR per 250g bag" or "Create a category called Espresso Blends."
• Promotions: "Create a 10% discount code WELCOME10" or "Set up a coupon for free shipping over 60 EUR."
• Content: "Write a blog post about how we source our beans" or "Generate descriptions for my products that are missing one."
• Settings: "Change the store currency to EUR" or "Set the opening hours to Tuesday through Sunday."

Keep design edits and catalog actions in separate prompts. One clear instruction per message gives the AI an unambiguous target and a clean commit.

Reference a look without naming a brand

If you want the site to echo an existing site's look, give the AI the public URL and ask it to match the style: "Make the layout feel like example.com." The AI fetches that page and extracts its palette, fonts, and section structure, then reuses those tokens when it builds. It reads the page over plain HTTP and uses only the visual signals; it does not copy content.

What good prompting gets you, and the privacy line

Your prompts and generated code are not sent to OpenAI or Anthropic. MaShop routes your request to a model by tier: short edits go to a fast model, multi-file features to a mid-tier model, full-site builds to the strongest model. You do not choose a vendor; you write a clear prompt and the router picks the right tier. See Models for how routing and tiers work, and How it works for the full generation loop.

Next: read The workspace to see where the chat, preview, and file tree live, and Versions and history to roll back a change you do not like.