Prompt Engineering for Beginners: 10 Mistakes and How to Fix Them (2026)

Prompt engineering for beginners mostly means fixing a few habits: asking in too few words, hiding the real goal, skipping the format, and trusting the first answer. These habits matter because a Kaspersky survey reported by The News found that 86% of professionals surveyed in Pakistan use AI tools for work, while 21% said they had received no AI-related training at all.

This guide walks through the ten prompt mistakes I correct most often in class. Each one comes with a weak prompt, a better version, and the reason the fix works. In the pages below you will find:

• The six parts that turn a vague request into a clear brief
• Ten common mistakes, each with before-and-after prompt examples
• A summary table you can keep next to your screen
• A simple loop for testing a prompt before you rely on it at work
• Tips for Urdu and Roman Urdu answers, and what never to paste into a chatbot

Why Do Beginners Get Weak Answers From AI Chatbots?

A chatbot only knows what you type plus what it learned in training. It doesn't know your shop, your manager, your class or your deadline. When a prompt leaves those out, the model fills the gaps with the most generic answer it can produce. That's why a short request often returns a polite, long and useless reply.

Anthropic's prompting documentation suggests a useful test: show your prompt to a colleague who knows nothing about the task. If that person would be confused, the model will be too. Its guide to clear and direct prompts compares the model to a capable new employee who hasn't learned your norms yet.

So most weak answers are not a sign that the tool is bad. They're a sign that the brief was thin, and good prompt engineering for beginners starts with fixing that brief. Learning how to write prompts is a skill you can practice in a week, and it works the same way in ChatGPT, Gemini, Copilot or Claude.

Prompt Engineering for Beginners: The Six Parts of a Strong Prompt

Different vendors use different labels, but they describe the same ingredients. Google's Gemini for Workspace prompting guide names persona, task, context and format. Microsoft's Copilot prompt guide uses goal, context, source and expectations. Put together, you get six parts:

1. Role: who the AI should act as, such as "an HR officer at a Lahore textile firm".
2. Goal: the exact result you want back.
3. Context: why you need it and who will read it.
4. Source: the facts, file or text it must work from.
5. Format: length, layout, tone and language.
6. Examples: samples of the output you like.

You rarely need all six. Microsoft's guide says a clear goal is the only required part, and you add the others when you want a more specific result. A one-line question about a word's meaning needs a goal and nothing else. A customer message that goes out under your shop's name needs most of the list.

Mistakes 1 to 3: A Request That Is Too Thin

Mistake 1: Asking in Five Words

Most weak ChatGPT prompts I see in class are one short line. Google's guide notes that the prompts people try first are often under nine words, while the prompts its users found most successful averaged around 21 words. Length is not the goal in itself. The extra words are where your details live.

Before: "Write a Facebook post for my shop."

After: "Write a Facebook post for a mobile accessories shop in Abbottabad. We're giving 15% off chargers and cables this Friday and Saturday. The readers are college students. Keep it under 60 words, friendly, and end with: WhatsApp us at [number]."

The second version tells the model what is on sale, for whom, for how long, and how long the post should be. There's very little left for it to guess.

Mistake 2: Hiding the Real Goal

Beginners often ask for an action ("summarize this") without saying what the result is for. Anthropic's documentation recommends explaining the reason behind an instruction, because the model can then make better choices on its own.

Before: "Summarize this report."

After: "I have a two-minute call with my manager tomorrow. From the report below, list the three decisions he needs to make, with one line of evidence for each."

Now the summary has a shape. It's short, it's ordered around decisions, and it skips the background your manager already knows.

Mistake 3: Forgetting the Reader

The same topic needs a different answer for a Class 10 student, a BSCS student and a bank manager. If you don't name the reader, the model writes for an average adult who may not exist.

Before: "Explain SQL joins."

After: "Explain SQL inner and left joins to a first-semester student who knows Excel but has never used a database. Use one example with a Students table and a Fees table. Keep it under 200 words."

Mistakes 4 to 6: Missing Format, Facts and Examples

Mistake 4: Not Saying What the Answer Should Look Like

If you don't specify a format, you'll usually get long paragraphs with headings you didn't ask for. State the length, the layout and the language in plain words.

Before: "Make a study plan for the computer operator test."

After: "Make a four-week study plan for a computer operator test as a table with three columns: Week, Topic, Daily Task. I can study 90 minutes a day. Keep the whole answer under 150 words."

A table is easy to check and easy to paste into Word or Excel. A word limit stops the model from padding.

Mistake 5: Asking the AI to Remember Facts You Should Supply

Fees, tax rates, admission dates and legal rules change often, and a chatbot's training data may be out of date. A 2025 research paper, Why Language Models Hallucinate, argues that common training and scoring methods reward guessing over admitting uncertainty. In practice, a model may give you a confident fee that is simply wrong.

The fix is to give it the source. Paste the notice, the price list or the policy text, then ask it to work only from that.

Before: "Write a fee reminder for parents with the current fee."

After: "Using only the fee schedule pasted below, write a polite SMS reminder for parents whose November fee is unpaid. Mention the due date from the schedule. If any detail is missing, write [check] instead of guessing."

For long documents, order matters too. Anthropic's guide advises putting the long text at the top and your question at the end. It reports that this ordering improved response quality by up to 30% in its tests.

Mistake 6: Describing a Style Instead of Showing It

Words like "professional", "catchy" or "simple" mean different things to different people. An example removes the guesswork. Anthropic's documentation recommends three to five examples that closely match your real task and differ enough that the model doesn't copy one pattern too strictly.

Before: "Write a Daraz product description in a good style."

After: "Here are three product descriptions from our store that sold well: [paste]. Write one in the same structure and tone for a steel water bottle, 750 ml, keeps water cold for 12 hours."

Mistakes 7 to 10: Trusting, Overloading and Oversharing

Mistake 7: Stacking Several Jobs Into One Prompt

"Read my CV, fix it, write a cover letter, and suggest five jobs" is four tasks. The model will do each one halfway. Anthropic's guide suggests numbered steps when the order or completeness of a task matters, and separate prompts work even better for big jobs.

After: Send one message per job. First: "List the five biggest problems in this CV for a junior accounts role." Next: "Rewrite only the Experience section to fix problems 1 and 2." The cover letter comes after you're happy with the CV.

Mistake 8: Accepting the First Answer

Both Google and Microsoft describe prompting as a back-and-forth. Microsoft also points out that the same prompt can return slightly different answers each time. Treat the first reply as a draft, then steer it with short follow-ups:

• "Shorter. Keep only points 1 and 3."
• "Make the tone more formal; this goes to a bank."
• "Rewrite the second paragraph for someone who has never used Excel."

Mistake 9: Copying Facts Without Checking Them

Microsoft's guide tells users to evaluate responses and cross-check them with trusted sources. This matters most for numbers, names, laws, medical advice and anything you'll sign. A quick habit helps: ask the model to mark any claim it is unsure about, then verify those claims on an official website before you use them.

Mistake 10: Pasting Private Data

Beginners paste whole customer lists, CNIC numbers, salary sheets and even passwords into free chatbots. Don't. Check your employer's AI policy first. When you need help with a real document, replace private details with placeholders such as [Customer 1], [CNIC] or [Amount]. The model can still fix the wording, and nothing sensitive leaves your computer.

All 10 Mistakes in One Table

Keep this table open while you practice prompt engineering for beginners. If you only change five habits this week, start with the swaps below. They cover the problems I see most often in student prompts.

How Do You Test a Prompt Before Using It at Work?

A prompt you use once can be fixed on the spot. AI prompts for work are different: one you'll reuse every day, such as a prompt that drafts replies to customer enquiries, deserves a short test. Anthropic's prompt engineering overview says to start with clear success criteria, a way to test against them, and a first draft. Here's a simple version you can run in a notebook, with no coding needed:

1. Define success. Write three to five things a good answer must contain, for example "mentions the price, under 80 words, polite, no promises about delivery dates".
2. Draft the prompt. Use the six parts from earlier.
3. Test it on three real inputs. Pick different cases: an easy enquiry, an angry one and a confusing one.
4. Check each answer. Tick your list from step 1 for every reply.
5. Change one thing and test again. If you change three things at once, you won't know which change helped.

Save the final version in a document with a short note on when to use it. Over a month, you'll build a small library of tested prompts for your own job.

Prompt Writing Tips for Urdu and Roman Urdu Answers

Many customers in Pakistan write in Roman Urdu on WhatsApp, and a reply in stiff English can feel cold. Say the language and script directly in the prompt, because the model won't always guess.

• Name the script: "Reply in Roman Urdu, not Urdu script" or "Reply in Urdu script".
• Paste one or two of your own past replies so it copies your everyday wording.
• Ask for simple words: "Use words a shopkeeper in Mansehra would use."
• Check names, prices and dates yourself, since mistakes are easier to miss in a mixed-language reply.

Is Prompt Engineering a Separate Job in Pakistan?

For most people, no. In job ads you're more likely to see prompting as one skill inside roles such as content writer, digital marketer, virtual assistant, data analyst or automation specialist. Employers want someone who can get reliable output from AI and plug it into real work.

That's why prompting pairs well with a practical skill. Writers use it to draft faster, analysts use it to clean and explain data, and automation specialists put tested prompts inside workflows that run every day. If you're choosing what to learn next, our posts on AI skills for jobs and free AI tools and their limits cover the wider picture.

Practicing Prompts Inside PSTA's AI Automation Course

Prompt engineering for beginners is where our AI Automation Course starts, but it doesn't stop there. It's an 8-week program for beginners. You can attend classes at our Abbottabad campus, follow them live from home, or mix the two. The current fee is Rs. 22,000 (regular Rs. 28,000), payable in 2 installments.

The outline includes modules on building chatbots with large language models and connecting AI APIs, including OpenAI's, to tools such as n8n, Make and Zapier. That's where prompt habits pay off: a prompt inside a workflow runs hundreds of times, so a vague instruction turns into hundreds of weak replies. Projects include a lead follow-up workflow, an invoice reader and a WhatsApp support bot.

If you'd like to see how automation fits your plans first, read our 8-week AI automation roadmap. You can also browse all PSTA courses or check the steps on the admission page.

Tonight's Exercise: Rewrite One Prompt You Used This Week

Open your chat history and find one answer that disappointed you. Rewrite that prompt using the six parts: role, goal, context, source, format and examples. Run both versions and compare them side by side. You'll usually see the difference on the first try, and that one comparison teaches more than a long list of tips.