Prompting guide

The agent is capable, but how you talk to it makes a real difference. This guide covers the habits that get you faster, more accurate results, then shows real prompts rewritten from vague to sharp so you can see the difference in practice.

The six habits

1. Start a fresh chat for a new task

The agent works best with focused context. When you move on to something unrelated, start a new chat. A long thread that has drifted across many different topics is harder for the agent to reason about than a clean one aimed at a single goal.

2. Show, do not just tell

Words only go so far. When something looks wrong, show it:

• Attach a screenshot to your message. You can attach up to 4 images or documents per message, 8 MB each, and this is the single highest-leverage habit on this page.
• Point at the element with Edit Mode if the change is visual, and skip the agent entirely.
• Name the page and describe exactly where on it the problem is.

The agent fixes what it can see far more reliably than what it has to guess at.

3. Ask by outcome, not by implementation

Describe the result you want, not how you think it should be built. "Users should be able to reset their password from the login page" is better than a step-by-step guess at the code. The agent is good at figuring out the how once it understands the what.

4. One thing at a time

A focused request lands more reliably than five unrelated changes packed into one message. Send the first change, watch it land, send the next. The loop is fast enough that small steps beat big batches. (Several small changes to the same thing are fine to batch: "on the pricing cards, make the price bigger, add a badge to the middle card, and gray out the annual toggle" is one focused request.)

5. Plan the big things before building them

For a large feature, use Plan mode first. The agent will read your project, ask you clarifying questions, and produce a written plan with the goal, scope, and open questions resolved, without touching your app. Then switch to Build and point at the plan. Ambitious features go much better when the thinking happens before the building. Plan mode is a paid feature.

6. Expect a few misses, and correct quickly

AI is not human. It is a statistical model, and it will always make some mistakes. That is not a sign anything is broken. When a change is not right, tell the agent what is off and it will adjust. Short, specific feedback beats a long re-explanation, and auto-fix already catches a whole class of problems before they reach you.

Prompts, rewritten

The fastest way to internalize the habits is to see them applied. Each pair below is a real kind of request, first as commonly written, then rewritten the way the agent hears best.

Starting a project

Instead of:

Make me a website for my business.

Try:

Build a site for a dog grooming studio called Bark and Bubbles. Home page with photos and a "book now" button, a services page with three packages and prices, and a booking page where customers pick a service, a date, and a time slot, and leave their name, phone, and email. Warm and friendly, cream and brown colors.

Why it works: the agent gets the name, the pages, the data each page needs, and the feel. Every one of those would otherwise be a guess.

Changing a design

Instead of:

The homepage looks bad, make it more modern.

Try:

On the homepage: make the hero photo full width, move the headline on top of it in large white text, and reduce the three feature cards below to one row with smaller icons. Keep the colors as they are.

Why it works: "modern" means something different to everyone. Naming the elements and the change to each removes the guesswork, and saying what to keep protects the parts you like. For changes like these, also consider Edit Mode, which skips the agent entirely.

Fixing a bug

Instead of:

The form is broken.

Try:

On the Contact page, when I fill in the form and press Send, the button shows a spinner forever and no message appears in the Messages table. Attached is a screenshot. It worked before you added the newsletter checkbox.

Why it works: where it happens, what you did, what you expected, what happened instead, and when it last worked. That last detail is often the whole diagnosis.

Adding a feature with rules

Instead of:

Add discounts.

Try:

Add a discount code field to checkout. Codes live in a table I can edit, each with a percentage and an expiry date. An expired or unknown code shows "This code is not valid" under the field. A valid code shows the new total immediately. Only one code per order.

Why it works: business rules are the number one place where the agent's guess and your intent quietly differ. Spelling out the edge cases (expired, unknown, stacking) gets them built correctly the first time instead of discovered later.

Working with data

Instead of:

Users should have profiles.

Try:

Members can fill out a profile: photo, display name, bio up to 200 characters, and city. Profiles are visible to other signed-in members on a Members page, but the email address never is. Members can edit only their own profile.

Why it works: it names the fields, the length rule, and, most importantly, who can see what. Visibility is a privacy decision, and stating it up front means the agent declares the table correctly instead of defaulting.

Giving feedback on a miss

Instead of:

That's not what I meant, undo all of this.

Try:

Close, but two things: the badge should sit on the image's top right corner, not below it, and the price should stay visible when the card is hovered. The rest is right.

Why it works: it keeps the 80 percent that landed and names only the deltas. (If the change is wholly wrong, the Undo button restores everything at no token cost, and you can rephrase from a clean slate.)

Where you come in

The agent handles the how; the habits above are about supplying the what. The things only you can supply are exact wording and content, business rules and their edge cases, who can see which data, and what "done" looks like. A prompt that carries those four things almost always lands in one attempt, which is also the cheapest way to spend your tokens.