Squeeze My Prompt
All tasks

I want to build an app.

Every question is optional. The ones marked Recommended matter most.

For example: A shared grocery list for my household, so we stop buying duplicates.

For example: Two adults, on their phones, in the supermarket aisle, in a hurry.

For example: 1. Add an item. 2. Tick it off. 3. See items others added, live.

For example: Phone A adds 'milk'; phone B sees it within 2 seconds.

For example: Calm and friendly, like a paper notebook. Warm off-white, one green accent.

For example: No login yet; I want to test the idea first.

For example: Lists and items. Only people in my household can see our list.

For example: Phone first, in the browser; no app store.

For example: Existing Next.js app; follow the pattern in /components/List.

For example: Not a programmer. Explain how to run it.

For example: Empty list shows a friendly hint; offline adds sync later.

For example: Send a weekly email summary.

For example: React and Supabase, because I already know Supabase.

For example: Stores emails; other households must not see them.

More detail, if you have it

How should the AI start?
Saved on this device only

How to brief an AI to build an app

Use this form when you want an AI tool, such as Lovable, Bolt, v0, Replit or Cursor, to build something with screens, data and actions. For a site people mainly read, use the website form.

  • The idea. Every tool maker’s guide starts with what you are building and the problem it solves. Lovable’s prompting guide asks for the product, its audience and its purpose first.
  • Who uses it, and when. In tests published in Vercel’s guide to prompting v0, a vague prompt without this context produced a layout that did not adapt to phones, and a search and cart that did not work.
  • What it must do. Listing the main actions, most important first, helps keep the AI from inventing features or missing ones (Vercel). GitHub’s spec-kit template builds a specification around user stories in priority order, each testable on its own.
  • How you’ll know it works. Anthropic’s guide to Claude Code says that when the AI reports work as done and it is not quite right, the usual cause is that it had no way to check its own work. A test you would run gives it one.
  • Look and feel. Without a named style, AI tools fall back on a generic look. Anthropic and OpenAI both describe this and both advise naming an aesthetic.

If your app stores anything about people, open “Data and who sees it”. In Veracode’s 2025 tests of more than 100 models on 80 coding tasks, 45% of the code contained a well-known type of security flaw.

A worked example

A household wants a shared grocery list. Their answers: the idea, two adults shopping on their phones in a hurry, three things it must do, a two-phone test, a calm notebook look, and no login yet. The form composes:

I want to build this: A shared grocery list for my household, so we stop buying duplicates.

<context>
Who uses it, and when: Two adults, on their phones, in the supermarket aisle, in a hurry.
Starting point: Starting fresh.
</context>

<task>
What it must do, most important first: 1. Add an item. 2. Tick it off. 3. See items others added, live.
Done means: Phone A adds 'milk'; phone B sees it within 2 seconds.
Build the first version described here.
</task>

<constraints>
Look and feel: Calm and friendly, like a paper notebook. Warm off-white, one green accent.
Not in this version: No login yet; I want to test the idea first.
</constraints>

Plan first. List any assumptions rather than guessing. If the information here is not enough to answer well, say so and tell me what you would need. If something missing would materially change the result, ask me up to 3 short questions first. Otherwise proceed and state your assumptions.

What the form adds for you

It asks the AI to plan first and list its assumptions, because Cursor’s guide calls planning “the most impactful change”.