Good ChatGPT prompts share one habit: they say who the answer is for, what the task is, what material to use, and what shape the result should take. Vague requests get generic results, and specific requests get something you can use or correct quickly. This guide gives a simple framework for writing ChatGPT prompts, before and after examples for writing, study, coding and business, the mistakes that waste the most time, and a checklist for catching errors in whatever the model returns. All examples below are prompts only. No model replies are shown, because output varies from one run to the next.
What makes a ChatGPT prompt work?
A prompt works when it removes guesswork: the model knows the goal, the audience, the inputs and the format. OpenAI’s own prompt engineering guide advises being explicit, separating instructions from supporting material with clear delimiters, and adding a few input and output examples when you need a particular pattern.
The same guide states that generated content is non-deterministic. In practice this means two runs of the same prompt can differ, so treat the first reply as a draft. Prompting is a loop: ask, read, adjust the instruction, ask again. Features and model names change often, so check the current ChatGPT help pages for what your plan includes.
What is a simple framework for writing prompts?
Use five parts: role, task, context, constraints and format. You do not need all five every time, but the more of them you supply, the less the model has to invent.
- Role: who the model should act as, such as a copy editor or a patient tutor. This sets tone and level.
- Task: one clear action verb, such as rewrite, explain, summarize, compare or debug.
- Context: the facts, text or data the answer should rest on, plus the audience.
- Constraints: length, things to avoid, reading level, what to do when information is missing.
- Format: a table, a numbered list, plain paragraphs, or a specific structure.
A template you can reuse:
Act as [role]. [Task] for [audience]. Use only the material between the triple quotes. Constraints: [limits]. If something is missing, say so instead of guessing. Format: [format]. “””[your text]”””
The delimiter and the “say so instead of guessing” line are small additions with large effects. The first keeps your pasted text separate from your instructions, and the second gives the model a safe way out when it lacks information.
What do before and after prompts look like for writing?
For writing, the fix is to name the audience, the length, the voice and the source material. A bare request produces average text, and a detailed brief produces text you can edit rather than rewrite.
Before:
Write a blog post about remote work.
After:
Act as an editor for a practical business blog. Draft an outline for a 900 word article for team leads at small companies who are moving from office to remote work. Include five H2 headings phrased as questions, two sentences describing what each section will cover, and one concrete example per section. Avoid buzzwords and avoid claims about statistics. Format as a numbered list.
Notice that the second prompt asks for an outline first. Splitting a large job into stages (outline, then sections, then edit) usually gives better control than requesting a finished piece at once. When you paste a draft for editing, ask for tracked suggestions rather than a silent rewrite, so you can see what changed. After you get a revision, the Text Diff Checker compares two versions side by side once you press its compare button, which helps you spot edits that quietly altered your meaning.
What do before and after prompts look like for study?
For study, state your current level, the topic, and the kind of help you want, such as an explanation, practice questions or a quiz. Asking the model to teach in steps and check your understanding works better than asking for a summary.
Before:
Explain photosynthesis.
After:
Act as a patient tutor for a 15 year old studying biology. Explain photosynthesis in three short stages, using one everyday analogy. Then ask me three questions, one at a time, and wait for my answer before giving feedback. If my answer is wrong, explain the idea differently rather than repeating yourself. Do not give the answers to the questions in advance.
Two habits protect your learning. First, ask for the reasoning, not only the result, so you can follow it. Second, check facts against your textbook or class notes, because a fluent explanation can still contain an error. If you are preparing a written assignment with a word limit, the Word Counter counts words in your pasted text after you press its button, so you can see whether a generated draft fits before you edit it down.
What do before and after prompts look like for coding?
For coding, give the language and version, the exact error message, the relevant snippet, and what you expected to happen. Code problems are specific, so the prompt should be too.
Before:
My Python code doesn’t work, fix it.
After:
I am using Python 3.11. The function below should return the average of a list but raises ZeroDivisionError when the list is empty. Here is the code between triple quotes and the full error message below it. Explain the cause in two sentences, then show a corrected version that returns None for an empty list. Keep the function signature unchanged and do not add new libraries. Finally, list three test cases I should run, including the empty list.
Never paste secrets such as passwords, API keys or customer data into a prompt. Replace them with placeholders first. When the model returns code, run it yourself and run tests, because code that reads well can still fail. If you ask for a change to a larger file, a diff view makes review easier, and the Text Diff Checker is useful here too for comparing your original against the suggested version.
What do before and after prompts look like for business?
For business tasks, supply the audience, the goal, the facts you can vouch for, and the tone. Never ask the model to invent numbers or customer details you have not provided.
Before:
Write an email to my customers about the price change.
After:
Act as a customer communications writer for a small online store. Write a 120 word email telling existing customers that our subscription price rises from the old price to the new price on the effective date. Use the facts below only: [list your facts]. Tone: direct, polite, no exaggeration. Explain the reason in one sentence, tell them how to cancel, and give a clear subject line. Offer two subject line options. Do not promise anything that is not in the facts.
Business text often has hard limits, such as subject lines, ad headlines or product titles. A prompt can state them, but models do not always count precisely, so verify. The Character Counter shows the exact character count of any text after you press its button. If the model hands back text in the wrong capitalization for a headline, the Case Converter switches it between upper case, lower case and title style. For other quick text jobs, browse the free tools.
What are the most common prompt mistakes?
The most common mistakes are vagueness, asking for too much at once, and trusting the first reply. Each one is fixable with a small change in how you ask.
- Being vague. “Make it better” gives the model no target. Say better in what way: shorter, friendlier, more formal, simpler.
- Stacking many tasks. A prompt with ten requests tends to handle some well and others badly. Split the work into steps.
- Leaving out the source. If you want an answer based on a document, paste the relevant text. Otherwise the model answers from general patterns, which may not match your facts.
- Asking for facts it cannot know. Recent events, private data and exact figures are risky. Provide the data or verify it elsewhere.
- Forgetting the format. If you need a table, ask for a table, and name the columns.
- Accepting the first draft. Follow up with specific edits such as “shorten the second paragraph” rather than regenerating from scratch.
- Pasting sensitive information. Check your plan’s data settings and avoid including confidential material.
- Trusting confident tone. Fluent writing is not evidence of accuracy.
How do you check ChatGPT output for errors?
Treat every output as a draft from an assistant that can be wrong, and check it in layers: facts, logic, instructions followed, and wording. The more costly a mistake would be, the more checking it deserves.
- Check factual claims at the source. Names, dates, figures, quotes and citations should be confirmed on an official or primary page. If a reference looks specific but you cannot find it, assume it may be invented.
- Check against your own material. If you supplied text, make sure the answer does not add claims that were not in it.
- Check the constraints. Did it respect the length, the format and the things you told it to avoid? Count words and characters rather than trusting the model’s own count.
- Compare versions. When asking for an edit, compare before and after to find unrequested changes to numbers, names or meaning.
- Test code. Run it with normal input, empty input and unusual input.
- Ask for uncertainty. You can ask the model to list which statements it is least sure about, then verify those first. This helps prioritize, but it is not a substitute for checking.
- Get a human review. For legal, medical, financial or safety topics, use ChatGPT only for general background and rely on qualified professionals for decisions.
FAQ
Does a longer prompt always give a better answer?
No. Length helps only when it adds useful detail. A short prompt with a clear goal, audience and format often beats a long, rambling one. Remove anything that does not change what you want.
Should I give ChatGPT a role?
It can help set tone and level, such as asking for a tutor or an editor. It does not give the model real expertise, so pair the role with specific instructions and still check the result.
Why do I get different answers to the same prompt?
Generated content is non-deterministic, as OpenAI’s documentation notes. If you need a consistent structure, specify the format in detail and consider including an example.
How many examples should I include in a prompt?
A few is a common starting point. Provide two or three input and output pairs that cover different cases, then adjust. Too many examples can make the output copy their wording.
Can I trust ChatGPT’s sources and quotes?
Not without checking. Open each cited source yourself and confirm the claim appears there. If you cannot verify a quote or reference, leave it out.
How do I keep prompts I like?
Save them in a notes file as templates with placeholders in square brackets. Note what worked and what you changed, so you can reuse and improve them over time.
What should you take away from this?
Write prompts like a brief for a capable but literal colleague: clear goal, supplied material, defined format, and room to say “I don’t know”. Then review the result with the same care you would give any draft. Small habits, such as splitting tasks and comparing versions, matter more than any clever wording.
Free Tools Mentioned in This Guide
See all of them on the free online tools page.