Write prompts that work

Learn how to adjust your wording to get dramatically better results from the agent - without needing to know any code.

The anatomy of a strong prompt

A strong first prompt includes four elements (from the Your first build guide):

1. Purpose - what the app does in one sentence
2. Core features - 3-5 bullet points of what users can do
3. Design hint - visual style or mood (e.g. "clean and minimal," "warm and colorful")
4. How it should feel - the experience of using it (e.g. "fast data entry," "one-tap actions")

You don't need to specify tech stack, database, or file structure - the agent handles those details. Focus on what the app should do and how it should feel.

Start rough, shape it in Plan mode

You don't need a perfect request to begin. Switch the composer's Build / Plan switcher to Plan (or press ⌥P) and send your rough idea. The agent responds with a plan proposal card, with two options: Request changes or Looks good, let's build it. Plan mode has the full context of your app, so it shapes a better plan than trying to write the whole request in one shot. Nothing is built until you approve.

Plan mode is unavailable in Brainstorm, and E3 sessions.

Ready-made starter prompts

Here are a few examples you can copy and adapt:

Simple booking app:

"Build an appointment booking app. Users can see available time slots for the next two weeks, book a slot with their name and email, and get a confirmation. Use a calendar view that feels clean and minimal."

Small online store:

"Create a small store where I can list products with photos, prices, and descriptions. Customers can add items to a cart and check out (no payment for now - just collect their info). Keep the design warm and colorful, with big product images."

Personal portfolio:

"Make a portfolio site where I can show my projects. Each project has a title, image, description, and link. Include an about-me page and a contact form. The design should feel clean and professional, easy to skim."

Before and after: fixing vs adding

When something doesn't work the way you want, describe the problem clearly rather than guessing at a solution.

Instead of:

"Change the button color to blue"

Try:

"The submit button on the bottom right is hard to see against the background. Make it stand out more."

This gives the agent room to fix the root issue (contrast) rather than just applying your guess (blue might not solve it).

Instead of:

"Add a filter dropdown"

Try:

"Let me filter books by read/unread status. A toggle or tabs would work."

The second version states the goal and suggests patterns for how your app looks and feels, but doesn't lock the agent into one approach.

Use specific examples

When you want a particular behavior, show a real example.

Vague:

"Make the form easier to use"

Specific:

"Right now I have to tap five fields to add a book. Can we use a floating action button that opens an inline form, like in the Google Keep app?"

The agent can infer patterns from familiar reference points (app names, conventions for how apps look and feel, or even "like how Stripe does checkout").

Keep long projects on the rails

Two habits that prevent the agent from re-breaking earlier work on a long-running app:

• Ask the agent to maintain a changelog of features and fixes built so far, it gives the agent a reliable record of what it has already done as the conversation grows.
• Scope big cleanups down. Broad prompts like "clean up / reorganize everything" invite sweeping changes and regressions. For refactors, ask for targeted, incremental edits, and consider forking first as a safety net.

If something looks wrong

If the agent misunderstands your prompt, reply with clarification in the same chat thread. You don't need to start over - just say "Actually, I meant..." and describe the gap.

The agent keeps context from earlier messages, so follow-up corrections are faster than rewriting the entire prompt.