A good ChatGPT prompt covers five things: the goal, the context, the task, the format and a check for gaps. Most weak answers come from leaving one of those out, not from a lack of clever tricks. That is all "prompt engineering" means for an everyday user, and you can learn it in one sitting.
This guide is for students, job seekers, small business owners and beginner coders in Nigeria who already have ChatGPT, Claude or Gemini on their phone and keep getting vague answers. Our prompt library, The Best ChatGPT Prompts for Work in 2026, gives you ready-made prompts and its own role-context-task-format formula. This guide differs in three ways: it adds a Check step for gaps and invented facts, it shows Nigerian before-and-after rewrites, and it compares what OpenAI, Anthropic and Google each say in their own guidance.
A prompt is a brief. The more a stranger would need to guess to do your job, the more the AI will guess for you.
Why most prompts fail
AI chat tools answer what you typed, not what you meant. Google's own prompting guide for Gemini tells users to "Write as if you're speaking to another person" and to "Express complete thoughts in full sentences", and to give as much context as possible (see Google's guide to writing effective prompts). Anthropic puts the same idea more bluntly: think of the model as a brilliant new employee who lacks context on your norms and workflows, so the more precisely you explain what you want, the better the result (Anthropic's prompting best practices).
Weak prompts tend to fail in five predictable ways, and our prompt library covers the basic weak-versus-strong comparison. The ones that matter most here:
- Too short. "Write my CV" gives the tool nothing about you, the job or the country.
- No audience. An email to a bank manager and a message to a friend need different words.
- No source material. Without your notes or text, the tool fills gaps from general knowledge, which is where made-up details come from.
- No format. If you do not say "five bullet points" or "under 120 words", you get whatever length the tool prefers.
- Too many jobs in one message. Google's developer guide recommends breaking complex tasks into separate prompts and chaining the steps (Gemini API prompting strategies).
A simple structure: Goal, Context, Task, Format, Check
You do not need a template for every message. For anything that matters, run through these five parts. Skip any part that does not apply.
- Goal. What is this for, and who will read it? ("A cover letter for a graduate trainee role at a bank in Lagos.")
- Context. The facts the tool cannot know: your background, the job text, your notes, the customer's message. Paste them in.
- Task. One clear action: draft, rewrite, explain, compare, summarise, check.
- Format. Length, tone, layout. ("Under 250 words, plain English, three short paragraphs.")
- Check. Tell it how to handle gaps. ("If something is missing, ask me before you write. Do not invent facts.")
This is our own simplification of the advice in the vendors' guides. Google's Workspace guide names four parts (persona, task, context, format) and says you do not need all of them every time. OpenAI's developer guide describes a similar order for apps: identity, instructions, examples, then context (OpenAI prompt engineering guide). The Check step is ours, and it is the one beginners skip most.
Before and after: five rewrites
These examples are illustrative. We wrote them to show the method and we have not published the tools' outputs here, so do not treat any sample as a guaranteed result. Run them yourself and compare.
1. CV and cover letter
Before: "Write a cover letter for me."
After: "I am a 2025 Computer Science graduate from a Nigerian university applying for a graduate trainee role in a bank. Below is the job advert and my CV. Write a cover letter of under 250 words in plain, confident English. Use only facts from my CV. Mention two skills from the advert that match my experience. If the advert asks for something my CV does not show, tell me instead of inventing it. [paste advert] [paste CV]"
What changed: goal, source material, length, tone and the Check step. The tool now has less room to invent a degree class or an internship you never did.
2. Studying
Before: "Explain photosynthesis."
After: "I am preparing for a school exam and I get confused by the two stages of photosynthesis. Explain them in simple English with one everyday comparison. Then ask me three short questions, one at a time, and wait for my answer before giving the next. Tell me which of my answers were wrong and why."
What changed: you stated your level and your confusion, and you turned a lecture into practice. Cross-check anything exam-critical against your textbook or teacher.
3. Small business content
Before: "Write a post for my shop."
After: "I sell secondhand phones in Ikeja. Write three WhatsApp status captions for a weekend promo on phones with a 30-day warranty. Friendly Nigerian English, no more than 25 words each, no emojis, and do not mention any price. I will add the price myself."
What changed: the product, place, offer, tone, limit and an explicit "do not". Leaving out prices on purpose avoids the tool inventing numbers.
4. WhatsApp customer replies
Before: "Reply to this customer."
After: "You are helping me answer customers for my bakery. Below is a customer message complaining that her cake arrived 40 minutes late. Draft a reply in two or three sentences: apologise once, state what we will do (a 10 percent discount on her next order), and keep the tone warm and respectful. Do not promise a refund. [paste message with the customer's name and phone number removed]"
What changed: the policy lives in your prompt, not in the tool's imagination. Removing the name and phone number is part of the method, and the privacy section below explains why.
5. Code help
Before: "My code is not working."
After: "I am a beginner learning Python. This function should return the average of a list of numbers but it returns the wrong number for [4, 6, 8]. Here is the code and the exact error or output. Explain the bug in plain language first, then show the corrected version, then tell me what I should learn so I can spot this myself next time. [paste code]"
What changed: expected behaviour, actual behaviour, the code itself and a request to teach you rather than just patch it. If you are starting from zero, our guide on how to learn to code in 2026 gives a path to follow.
Show it what you want with an example
When you care about style, one short sample beats a paragraph of description. Paste a reply you wrote before and say "match this tone". Anthropic describes examples as one of the most reliable ways to steer format, tone and structure, and Google's developer guide recommends a few varied examples while warning against too many, because the model can overfit its answer to them. OpenAI's guide calls this few-shot prompting. For a small business, this is the easiest upgrade: keep two or three of your best customer replies in a note and paste them in whenever you need a new one.
Common mistakes
The basics (vague prompts, no format, one giant request) are covered in our prompt library. Three that catch beginners out:
- Treating the first answer as final. Reply with "shorter", "more formal" or "use my notes only".
- Telling it what not to do without saying what to do. Anthropic recommends stating what to do instead. "Write in plain paragraphs" works better than "do not use bullet points".
- Copying the answer straight to a client, lecturer or employer. You are responsible for what you send.
How to check and fix a wrong answer
AI tools can state wrong things in a confident tone. Google's own Workspace guide says to review every output for accuracy and that the final output is your responsibility. A quick routine:
- Look for specifics. Names, dates, amounts, laws, quotes and links are the usual places for errors. Verify each one on an official page.
- Ask for the basis. "Which part of my notes did you use for this?" or "What are you unsure about?" can expose guesses, though the answer is not proof either.
- Ask for a rewrite with limits. "Rewrite using only the text I pasted. If a fact is not in it, write 'not provided'."
- Test code. Run it with a sample input before trusting it. Never run code you do not understand against real data or a live system.
- Start a fresh chat if it drifts. A long conversation can pile up wrong assumptions. Open a new chat and paste in only what is correct.
For recent facts, Google's developer docs suggest turning on search grounding where it is available. In consumer apps, look for a web search or sources option, and read the sources it lists yourself.
How ChatGPT, Claude and Gemini differ
The core method works everywhere. The vendors' own guidance differs mainly in emphasis. Note that OpenAI's and Anthropic's long guides are written for developers using the API, and the app on your phone may behave differently, so treat this as a guide to habits, not a spec.
| Tool | What its official guidance stresses | What to try |
|---|---|---|
| ChatGPT (OpenAI) | Standard GPT models need explicit instructions, while reasoning models do well with a high level goal. Use headers, lists or tags to separate instructions from pasted material. | Be specific and label your pasted text ("CV:", "Job advert:"). For hard problems, state the goal and let a reasoning model work out the steps, if your plan offers one. |
| Claude (Anthropic) | Be clear and direct, explain why you want something, say what to do rather than what to avoid, and wrap different content in XML style tags. | Add one line of purpose ("this is for a job application"). Use tags like <cv> and <advert> when you paste several documents. |
| Gemini (Google) | Persona, task, context and format; natural sentences; iterate. The Gemini app also suggests starting with "Make this a power prompt:" to have it improve your draft. | Start short, then refine in follow ups. Use the power prompt trick when you are unsure how to ask. |
Two cautions. Model names, plan features and menus change often, so check what your own app offers. And each tool answers differently, so try your own prompt in two tools and compare.
Privacy
Anything you type goes to a company's servers, so leave out passwords, OTPs, card numbers, BVN, NIN and customers' names and phone numbers. Our prompt library has a fuller list of what not to paste. OpenAI's help centre documents a data controls setting and a Temporary Chat option (OpenAI data controls FAQ). Confirm the options and defaults in your own settings. Claude and Gemini have their own privacy settings too.
FAQ
Do I need to learn "prompt engineering" to use ChatGPT?
No. For daily use, the five parts above are enough. Formal prompt engineering is a developer skill for building apps, and the vendors' docs treat it that way.
Should I start with "Act as an expert"?
It can help set tone, and Google's guide lists a persona as one optional part. It does not replace context. A role with no facts still produces a generic answer.
Is a longer prompt always better?
No. Add what is relevant and cut the rest. Google's guide also advises being concise and avoiding jargon. Long and specific is fine, long and rambling is not.
Can I use these prompts on the free plan?
The method does not depend on a paid plan. Features such as file uploads, web search and reasoning models vary by plan and change often, so check your app.
Why does the same prompt give different answers each time?
These tools generate answers with some variation, so exact repeats are not guaranteed. If a result matters, save the version you checked.
Takeaway
Write your prompt like a brief for a capable stranger: goal, context, task, format and a check. Treat every answer as a draft you are accountable for, and keep personal data out of the chat. For ready made starting points, see The Best ChatGPT Prompts for Work in 2026.
Last reviewed 23 October 2026. Tested: no. Nothing here was run in ChatGPT, Claude or Gemini. The examples are illustrative and were not run for this article. Vendor guidance changes, so check the linked official pages for updates.
