How to Write Good Prompts: A Practical Guide

Why prompts make such a difference

Large language models predict a likely continuation of the text you give them. When the input is short and vague, the most likely continuation is something average: a safe, general answer that fits thousands of situations and therefore fits yours poorly. When the input is specific, the model has far fewer plausible directions to go in, and the result fits your situation.

Think of briefing a capable freelancer who has never met you. They are skilled, fast and willing, but they know nothing about your company, your reader or your deadline. Whatever you leave out, they have to assume. A good prompt simply removes the assumptions that matter.

The seven building blocks below are not a rigid template. Short tasks need only two or three of them. But when an answer disappoints you, the cause is almost always one of these blocks missing or unclear.

1. Role: tell the AI who to be

A role sets the perspective, vocabulary and standards the model applies. "You are an experienced employment lawyer" leads to different word choice, caution and structure than "You are a friendly HR assistant", even for the same question.

Good roles are specific and relevant to the task:

• Weak: You are a helpful assistant.
• Better: You are a senior e-commerce copywriter who writes for outdoor equipment brands.

A role is not magic. It will not give a model knowledge it does not have, and it does not make the answer more reliable on its own. Its main value is steering tone and focus. Keep it to one sentence and spend your effort on context instead.

2. Context: give it what it cannot know

Context is the most underrated part of a prompt and usually the one that improves results most. The model does not know your product, your audience, what you tried before, what your boss said, or what happened in last week's meeting. If those facts matter, they must be in the prompt.

Useful context includes:

• Audience: who will read or use the result, and what they already know.
• Purpose: what the result should achieve, such as a booked call, a decision or fewer support tickets.
• Facts: names, numbers, dates, features, prices, constraints. These are the things you do not want the model to invent.
• Background: what has happened so far and any sensitivities.

When you paste longer material such as notes, a transcript or an article, separate it clearly from your instructions, for example with triple quotes or a heading like "Source text". This helps the model distinguish what to work on from what to do.

3. Task: one clear instruction

State the task with a verb and an object: write, summarise, outline, compare, rewrite, extract, classify. "Help me with my email" is not a task; "Write a follow-up email to a client who missed our call" is.

If a job has several stages, such as research, outline, draft and edit, you will usually get better results by doing them one at a time. Ask for an outline, review it, then ask for the draft. Each step gives you a checkpoint where you can correct direction before the model builds on a mistake.

It also helps to say what success looks like. "The reader should be able to set up the device without calling support" tells the model what to optimise for.

4. Constraints: set the boundaries

Constraints tell the model what not to do and how far to go. Common ones:

• Length: "under 150 words", "three paragraphs", "one sentence per bullet".
• Tone: "professional but warm", "plain English, no jargon".
• Accuracy: "Use only the facts above. Do not invent statistics, quotes or sources."
• Uncertainty: "If something important is missing, ask me up to three questions before you start."
• Exclusions: words, claims or topics to avoid.

The accuracy and uncertainty constraints are worth adding to almost every prompt. Language models can produce confident text that is not true. Telling the model to use only your facts and to ask when unsure does not eliminate errors, but it makes them less likely and easier to spot. You still need to check anything that matters.

Phrase constraints positively where you can. "Write in short sentences" usually works better than "don't write long sentences."

5. Format: describe the shape of the answer

If you know what you will do with the answer, say so, and describe the format. A table you can paste into a spreadsheet, a numbered list someone will follow, a subject line plus email body, valid JSON for a script. Without a format, the model chooses one, and it often adds introductions and summaries you do not need.

Be concrete: "Return a Markdown table with the columns Task, Owner, Due date" is better than "make it organised." For longer outputs, list the sections you expect in order.

6. Examples: show, don't just tell

An example of the output you want is often more effective than a paragraph of description. If you have a past email whose style you like, paste it and say "match the tone and length of this example." If you need data in a specific structure, show one filled-in row.

Two cautions. First, models tend to copy examples closely, including their topic and phrasing, so say which aspects to imitate (tone, length, structure) and which to ignore. Second, one or two good examples are usually enough; many examples make the prompt long and can over-constrain the result.

7. Iteration: treat the first answer as a draft

Even a great prompt rarely gives a perfect first answer. The fastest path is a targeted follow-up in the same conversation, not a brand-new prompt:

• "Shorter: cut it by about a third."
• "Lead with the price, then the benefits."
• "Option 2 is closest. Make it more personal and drop the last paragraph."
• "What assumptions did you make? List them."

When you find a prompt that works for a recurring task, save it. Next time, change only the parts that differ. That is exactly what the builders on this site do: each one is a tested structure for one job, where you fill in only what changes.

Bonus: prompts that persist

Everything above also applies to system prompts: the standing instructions behind a custom GPT, a Claude Project or a Gemini Gem. The difference is that a system prompt has to work for many conversations you cannot see in advance, so rules and limits matter more than task details.

• State the assistant's purpose and users in the first lines.
• Name the sources of truth, such as uploaded files, and say what to do when they do not contain the answer.
• Write concrete rules ("answers under 150 words unless asked") instead of values ("be concise and helpful").
• List what is out of scope and where to send people instead.

The system prompt generator turns these points into ready-to-paste instructions. For code, the same principles apply in the code review and debugging prompt: say what the code should do, paste the exact error, and ask for the smallest fix.

Before and after: one prompt, improved

Here is a typical first attempt:

Write a post about our new opening hours.

The model has to guess the business, the platform, the new hours, the audience and the tone, so the result will be generic. Here is the same request with the building blocks added:

You are a social media copywriter for a small neighbourhood bakery.

## Task
Write 3 short Facebook post options announcing our new opening hours.

## Context
- From 1 March we open at 6:30 instead of 7:30, Monday to Friday.
- Reason: many customers asked for coffee before their commute.
- Audience: local commuters and parents doing the school run.

## Requirements
- Tone: warm and friendly, no exclamation marks overload.
- Under 60 words per option.
- First line must state the change.
- Use only these facts. Do not invent offers or prices.

## Output format
Label the options Option 1, Option 2, Option 3.

Nothing in the second version is complicated. It simply answers, in advance, the questions a human copywriter would ask before starting.

A checklist for every prompt

1. Did I say who the AI should be, if the perspective matters?
2. Did I describe the audience and the purpose?
3. Did I include the facts the model cannot know?
4. Is the task a single, clear instruction with a verb?
5. Did I set length, tone and accuracy constraints?
6. Did I describe the output format?
7. Would an example make it clearer?
8. Am I ready to iterate with a short follow-up instead of starting over?

Finally, keep your own judgement in the loop. Check facts, names and numbers before you publish or send anything, and do not paste information into an AI tool that your organisation does not allow there.

Try it with a builder

Each builder on PromptsLaunchpad applies this structure to one specific workflow, from emails and meeting summaries to resumes, business plans, spreadsheet formulas and image prompts. Pick the one that matches your task:

Or use the general prompt builder for any other task, or browse all 18 builders.