Lapidia

Guide

How to write a good prompt (with examples)

When an AI's answer comes out generic, the request almost always did too. The AI fills in whatever you didn't say with the most likely option, and the most likely option is the generic one. A good prompt spells out what only you know. In practice, that fits into four parts.

The 4 parts of a good prompt

  1. 1. Goal

    What you want at the end, phrased as a task. "Help me with the onboarding" leaves the AI guessing; "propose a 3-screen onboarding flow" tells it what to deliver.

    Propose an onboarding flow of up to 3 screens for a personal organization app.

  2. 2. Context

    Who it's for, what the situation is and what already exists. The AI can't see your screen, your project or yesterday's conversation: whatever isn't in the request, it makes up.

    It's for people opening the app for the first time who don't know how to create a list yet.

  3. 3. Constraints

    What's off-limits, the boundaries and the tone. Constraints are what pull the answer out of generic territory.

    Welcoming tone, one primary action per screen, no sign-up before the second screen.

  4. 4. Response format

    How you want to get it back: a table, a list, code, a ready-to-send email, a word count. Leave it out and the AI picks the format for you.

    Present it as a table with the title, supporting copy and action for each screen.

Put together, the four lines above make a complete prompt. Not every request needs all four, but almost every vague request is missing at least one.

Before and after: three examples

The draft the way people actually type it, and the same request with the missing parts filled in.

Development

Before

make a login screen in react

After

Build a login screen in React and TypeScript.

Use email and password fields with validation and clear error messages.

Include loading states, keyboard navigation and a responsive layout.

Deliver the complete component, styled with Tailwind CSS.

What was added: context defined · clear requirements · delivery format.

Design

Before

help me with my app's onboarding

After

Propose an onboarding flow for a personal organization app.

Design a flow of up to 3 screens for someone using the app for the first time.

Define a title, supporting copy and one primary action per screen, in a welcoming tone.

Present the flow in a table and explain the goal of each step.

What was added: audience defined · scope set · delivery format.

Writing

Before

write an email introducing my product

After

Write an email introducing Lapidia, an app that improves prompts.

Address professionals who use AI at work and want to phrase requests more clearly.

Use a warm, direct tone, up to 120 words, with no exaggerated claims.

Include a short subject line, one concrete benefit and an invitation to try the app.

What was added: audience defined · tone of voice · clear goal.

Common mistakes

  • Pointing to something the AI can't see. "Fix the error above", "make it like that one": paste the error, the snippet or the example into the request itself.
  • Asking it to "make it better" without saying how. Better in what way? Shorter, more formal, more persuasive, less jargon? Say what counts as better.
  • Cramming everything into one request. Planning, writing and reviewing in one go gets you a mediocre answer on all three. Break big jobs into steps.
  • Forgetting who it's for. A text for your leadership team and one for end customers call for a different tone, length and vocabulary.
  • Not asking for a format. If you're going to paste the answer into a spreadsheet, an email or your code, ask for that format from the start.

Checklist before you hit send

  • Does the request start with an action verb (write, list, compare, fix, create)?
  • Does it say who it's for and in what situation?
  • Is everything it refers to pasted into it (the error, the snippet, the data)?
  • Are the limits clear (length, tone, what to avoid)?
  • Does it ask for a response format?

Prompts for ChatGPT, Claude or Cursor: does anything change?

The four parts apply to all of them. What changes is the context each one already has: in Cursor, the AI can see your code, so your request can name files and functions directly; in ChatGPT and Claude, it only sees the conversation, so anything outside it has to be pasted in. In any of them, asking for the response format saves you a round of back-and-forth.

Do it with a shortcut

Lapidia does this work right in the field you're typing in: it reads your draft, spots which of the four parts is missing, asks for what only you know and rewrites the full request, in ChatGPT, Claude, Cursor or any app. Your text isn't stored, and password fields are never read.

Download Lapidia