An MCP server gives an assistant tools; the prompt decides whether it uses them. We watched teams ask for the same page eight different ways and kept the prompts that worked on the first try.
Prompts that search instead of guess#
- “Find a split hero for a fintech product, dark, with a product screenshot. Use the Next.js version.”
- “Show me three pricing sections with a monthly/annual toggle for a B2B SaaS.”
- “Get the footer from the template we used for the homepage and match its spacing.”
Each names a section type, an industry or a style, and a stack. The server’s search reads exactly those words, so the assistant calls it instead of writing a layout from memory.
Prompts that keep your brand#
- “Read DESIGN.md first and use it for every section today.”
- “Use our tokens — never type a colour.”
- “If a section needs a colour we don’t have, ask me instead of choosing one.”
Prompts that finish the page#
- “Assemble the sections we approved into one page, in this order: hero, logos, features, pricing, FAQ, footer.”
- “Now check the page at 375px and fix anything that overflows.”
What does not work#
- “Make it look modern.” — no section, no search, a guess.
- “A landing page for my startup.” — too big; the assistant writes everything at once.
- Pasting a screenshot instead of DESIGN.md — the assistant guesses values from pixels.
The pattern is the same every time: the more your prompt reads like a search query, the more the assistant searches. And the more it searches, the less you fix.