Custom Prompt Builder

How to use this

Start with the task itself. Describe what you want your AI to do, as specifically as you can, then say where in your project it goes and what must not change. Next, pick the kind of task, which adds a matching rule to the prompt and a matching item to the answer checks, and say whether your AI can see your project files or you'll paste the code. Add a part of your site to match, an error message or a note about a failed attempt if any of those apply.

The prompt updates as you type. Read it back once, copy it and paste it into your AI chat or coding tool. When the AI answers, go through the list of what to check in its answer before you keep the change. Press "Show why each part is here" at any time to see what each part of the prompt is for, and press it again to hide the notes.

Why your own words lead the form

There's no single perfect prompt you can copy for every job. The shape of a good prompt carries over from task to task, but what goes in it is specific to your project every time, which is the point our guide to writing prompts AI coding agents understand makes at length. So the builder never fills in your task with template text. What you want, where and what must stay the same come from you, word for word.

That also means it can't rescue a vague description. Type "the checkout page" and "fix it", and the prompt says exactly that. What the labels do is make the gap obvious: a vague "What" like "fix it" in an otherwise structured prompt stands out far more than the same vagueness buried in a paragraph.

Why every prompt carries the same rules

Your words say what to do. The rules say how far the AI may go while doing it, and they are the part a quickly typed prompt almost always leaves out. Every prompt from this builder tells the AI to change only what the task needs, to ask before touching anything outside the area you named, not to add new packages or outside services without your agreement and to stop and report after a failed attempt.

That last rule matters most. An AI that tries fix after fix on its own can end up changing far more than the original problem needed. Our guide to breaking vibe-code loops explains that loop. A stop after the first failure keeps you in charge of the next step.

Each task type then adds one rule of its own. A bug fix asks for the cause in plain words before any fix. A new feature asks for the smallest working version first. A tidy-up must leave everything working and looking exactly as it did. A connection to an outside service keeps secret keys out of the code visitors can see (our guide explains why an API key can't go in the frontend). A style change touches only the look.

Why it asks whether your AI can see your files

This one question changes the prompt more than any choice of AI does. An AI website builder or an AI coding agent working in your project can open the files itself, so the prompt asks it to look at the files involved and report what it found before changing anything, then to list every file it changed. An AI chat that isn't connected to your project only knows what you paste into it. For that case the prompt says so plainly, leaves a marked space for your code, asks the AI to request another file rather than guess at it and asks for each changed part back in full, not with gaps like "rest of code here".

Why the answer checks matter as much as the prompt

A good prompt makes a good answer more likely, not certain. The list under the prompt turns each rule into something you can see in the reply: did the AI report before changing, stay inside your "Where", leave your "must not change" list alone and tell you how to test the result? You don't need to read code for any of these. If a check fails, say so in your next message and quote the rule it broke.

Why "what must not change" always appears, even blank

It's one of the three parts every prompt needs and the one most often skipped, so the prompt keeps a line for it even when the field is empty. The FAQ below covers what happens when you leave it out.

Why the optional context is worth adding

A part of your site to match gives the AI something concrete to copy, rather than guessing what "cleaner" or "consistent" means. A pasted error message is the exact failure, not your summary of it. A note on a previous attempt stops the AI from offering the same fix that already failed, because the prompt now says what went wrong. None is required, and each usually saves a follow-up message.

A worked example

Here is what the builder produces for a common bug. The inputs:

• What: show a thank-you message after someone sends the contact form. Right now the form sends, but nothing appears.
• Where: the contact form on the Contact page
• What must not change: the form's fields, its design and every other page
• Task type: fix something that's broken. Access: the AI can see the project files.

The prompt it builds:

You can read this project's files. Before you change anything, look at the files involved and tell me briefly what you found and what you plan to change.

Task type: fix something that's broken.
What I want: show a thank-you message after someone sends the contact form. Right now the form sends, but nothing appears.
Where: the contact form on the Contact page
What must not change: the form's fields, its design and every other page

Rules for this change:
1. Change only what this task needs, and list every file you changed.
2. If the task needs anything outside the "Where" above, stop and ask me first.
3. If your first attempt doesn't work, stop. Tell me what you tried and what happened, instead of trying new approaches on your own.
4. Don't add new packages, libraries or outside services unless I agree first.
5. Find the cause first and explain it in plain words. Then fix the cause, not only the symptom.

When you're done, tell me in plain words what you changed and how I can check that it works.

And what to check in the answer: a cause named in plain words (for example, the form sends but nothing tells the page to show a message), a file list that stays inside the contact form, the fields and design untouched and a way to test it that you then try yourself by sending the form once.

What this builder can't do for you

It structures your task and adds the rules. It doesn't know your project, so it can't tell whether your "Where" names the right part, and it can't promise what any AI does with the prompt. Read the prompt back before you send it, and check the answer before you keep the change.

Frequently asked questions

Related reading

• Website Prompt Generator: ready-made prompts for 10 common website fixes, when your problem is a known one such as a page title or alt text.
• Describing a Bug So Your AI Agent Can Actually Fix It: what to put in the "What" field when something's broken.
• What Is an AI Agent: why an AI that acts on your project needs clearer limits than one that only answers.
• PayFast Webhook Code Generator: for once your next prompt is about wiring up a South African payment gateway.
• POPIA Compliance Checker: for once your next prompt touches real personal information.