Effective Prompt Writing: Notes from Teaching AI to Young People

In short: The key to good results from generative AI is less about finding the right tool and more about writing the right request. An effective prompt states the role, context, task, expected output format and limits clearly; treats the first answer as a draft; and verifies the result. It is also the topic we spend the most time on in the generative AI training I run for young people aged 15–30 at Habitat Association.

At Habitat Association I run generative AI training for young people aged 15–30 from all kinds of educational and professional backgrounds. Sessions are one-off, sometimes in person and sometimes online. In this article I describe both how we design the sessions and the topic at their centre: effective prompt writing.

How do the sessions work?

• Participants: young people aged 15–30 from every walk of life — students and working people, the tech-savvy and those trying AI for the first time.
• Format: one-off sessions. The aim is that every participant leaves with something they can use in their own work the next day.
• Method: the sessions are interactive, with as much room for practice as for explanation; prompts are written together and the results assessed together.
• Tools: alongside chat assistants such as Claude, ChatGPT and Gemini, the pool we draw on in the sessions includes more than 150 AI tools. The goal is not to teach every tool one by one, but to build the ability to choose the right tool for the need.

Why prompt writing?

People often ask why, with so many tools, I put prompt writing at the centre of the training. My answer: tools change; the skill of asking good questions stays. The assistant you use today may be different a year from now, but someone who can say clearly what they want gets value faster from every new tool.

What makes an effective prompt?

We use a simple framework in the sessions. A good request answers these questions:

You don't need all of them every time, but when a result misses the mark, what's missing is usually one of these.

Before and after: same request, two different results

Weak request:

Make me a presentation.

The model has to guess the topic, audience, length and tone. The result is usually generic and shallow.

Strong request:

I'm giving a 10-minute talk to high school students. Topic: what to
watch out for when using generative AI for homework.
Draft 8 slides: a title and at most three bullet points per slide.
Keep the tone friendly but instructive. On the last slide, include
three check questions students can ask themselves.

Audience, length, topic, output format and tone are all clear. Instead of guessing, the model does the job.

The first answer is not the final answer

The habit I stress most in the sessions is keeping the conversation going. The first answer is a draft. The fastest way to improve it is feedback:

• "Explain the second point with a more concrete example."
• "Make the tone more formal and the sentences shorter."
• "Turn this into a table."
• "List anything in this answer that may be missing or wrong."

That last one is especially valuable: asking the model to critique its own answer surfaces mistakes that would otherwise slip through.

Why is verification essential?

Generative AI gives fluent, confident answers, but fluency is not accuracy. Models sometimes produce non-existent sources, wrong dates or incorrect figures. The rules we give in the sessions are simple:

1. If you'll use information in homework, work or a decision, check it against a primary source.
2. If you ask the model for sources, verify that each source exists and actually contains that information.
3. Take extra care with figures, dates, names and quotes.

What shouldn't you share?

What you type into AI tools may be stored, depending on the tool's data policy. So:

• Don't share personal or sensitive data such as ID numbers, passwords or bank details.
• Don't enter other people's personal information (a friend's health condition, a customer's contact details) without permission.
• If you use AI at work, learn your organisation's rules for AI use.

Which tool should you use?

General-purpose assistants such as Claude, ChatGPT and Gemini each have different strengths, and those differences keep shifting. My practical advice in the sessions: for important work, try the same request in more than one tool and compare the results. For specialised needs such as images, audio, slides or code, choosing a suitable tool from a wider pool is often more productive than forcing a general assistant.

A prompt checklist for yourself

• Did I explain who I am and the situation?
• Can I say exactly what I want in one sentence?
• Did I specify the length and format of the output?
• Did I write what I don't want?
• Did I improve the first answer with feedback?
• Did I verify the information I'll use?
• Did I type anything I shouldn't share?

Frequently asked questions

What is a prompt?

A prompt is the request or instruction you type into a generative AI tool. The clearer and more contextual it is, the more accurate the answer.

How do you write an effective prompt?

By stating the role, context, task, expected output format and limits clearly. Treating the first answer as a draft and improving it with feedback is also part of effective prompt writing.

Who is generative AI training for?

Basic training suits everyone. The sessions I run at Habitat Association are for young people aged 15–30 from all backgrounds and require no prior technical knowledge.

Which is better: ChatGPT, Claude or Gemini?

Each has different strengths, and those differences keep changing. For important work, trying the same request in more than one tool and comparing the results is the most reliable approach.

Can you trust information from AI?

Not without verifying it. Models can produce fluent but wrong answers; figures, dates, sources and quotes in particular should be checked against primary sources.

Is it safe to type personal information into AI tools?

Avoid sharing sensitive data such as ID numbers, passwords or bank details, and other people's personal information. What you type may be stored according to the tool's data policy.