How to Write Better ChatGPT Prompts: 5 Parts, Real Examples

The way to write better ChatGPT prompts is to fill in five parts: who is writing, who is reading and why, what shape the answer should take, an example or two, and what to avoid. If you only fix one thing, fix the format line, because it is the cheapest way to turn a vague answer into one you can use as written.

My read is that good prompting is delegation, not wordsmithing. People who get consistent results are the ones who would also write a decent brief for a human contractor. Every example below is original and drawn from the kinds of work people hand to ChatGPT daily. I deliberately skip model version numbers and plan pricing, since those change faster than any prompt habit.

How to write better ChatGPT prompts: what actually goes in?

OpenAI’s own prompting guidance in its Help Center says much the same thing in plainer terms: put the instruction first, separate any source text with delimiters, be specific about length, format and style, and show the output format with examples. Here is that advice turned into a five-slot checklist.

Part	What it answers	What you get without it
Role	Whose voice is this?	Tone drifts between formal and chatty
Context	Who reads it, and why does it exist?	Generic textbook output
Format	Table, bullets, word count?	A wall of text you must reformat
Examples	What does “good” look like?	A style that misses your taste
Constraints	What to avoid, how long, what not to invent	Padding and confident guesses

You will rarely need all five. A quick email needs context and constraints. A data cleanup job lives or dies on format and examples. The skill is leaving a slot empty on purpose rather than by accident.

What does a weak prompt look like next to a strong one?

The fastest way to feel the difference is to run both. These pairs are rewrites I have actually made.

Task	Weak prompt	Stronger prompt
Apology to a vendor	”Write an apology email"	"Email to a vendor’s procurement lead. Our payment ran 9 days late. State the cause in one sentence, no excuses, say the corrected payment date will be confirmed in writing. Under 120 words, warm but direct.”
Meeting notes	”Summarize this"	"From the notes below, list decisions, owners and deadlines in a table. Put anything discussed but not decided in a separate ‘open’ column. Do not add information that is not in the notes.”
Spreadsheet help	”How do I total this in Excel?"	"Column A is date, B is store, C is amount. I want monthly totals per store starting in F2. Excel for Microsoft 365, no pivot tables. Explain each argument in one line.”
Brainstorm	”Ideas for a bakery promotion"	"Two-person neighborhood bakery, slow Tuesday afternoons, budget of $200 a month. Give 5 promotions in a table: idea, expected effect, risk, first step.”

Notice what the strong versions share. The reader is specific, the output has a shape, and at least one line tells the model what not to do. The weak versions force ChatGPT to answer questions you never knew you had left open.

Which prompt templates work for everyday tasks?

Save these with square brackets and swap the details. Over a month, a handful of them will cover most of what you ask.

Task	Template
Email	”To: [role, relationship]. Goal: [one line]. Facts that must appear: [1 to 3]. Tone: [warm/formal]. Max [N] sentences. Leave [ ] for anything you do not know.”
Summary	”Summarize the [document type] for [reader] who has 30 seconds. Three bullets plus one next step. Paraphrase, add no new facts.”
Spreadsheet	”Columns: [A=…, B=…]. Goal: [..]. App/version: [..]. Give the formula and one line on each argument.”
Code	”Language/version: [..]. Current behavior: [..]. Expected: [..]. Full error text: [..]. Show only the lines to change and explain the cause in two sentences.”
Planning	”Audience: [..]. Limits: budget [..], people [..], time [..]. Give [N] options as a table (option, benefit, risk, first step). Drop any that break the limits.”

For code, paste the full error text. A paraphrase makes the model guess the most common cause, which is often the wrong one. When I ask for styling help, I paste the actual utility classes from the page, much like the snippets in this site’s Tailwind tips post, and ask which ones conflict. For troubleshooting-style questions, giving the symptoms in order, as in the Wi-Fi speed fix guide, beats a one-line complaint.

How do I use examples and constraints without being copied?

Give two examples that differ on purpose. When I want blog headings, the routine is:

1. Write two sample headings, one a question and one that contains a number.
2. Add “make the new ones structurally different from both.”
3. State a character limit and any banned words.
4. Ask for eight candidates, then pick two and ask for a polish pass.

Phrase constraints as positive substitutes. “Avoid jargon” is vague; “only words a high-school student would know” is testable. The same goes for length: “about 150 words” gets you closer than “short.”

How should I use custom instructions, projects and memory?

OpenAI describes custom instructions as a place to set your role, tone and preferred output formats, applied to new conversations until you change them. Memory is for details you ask it to remember, and OpenAI has said deleting a chat does not delete the memories saved from it; those are removed separately. Temporary chats do not use memory. Press coverage also described a project-only memory option introduced in August 2025, though availability may differ by device and plan, so check the current settings rather than trusting any article, including this one.

My split:

• Custom instructions for rules that never change: answer first, flag uncertainty, no flattery.
• Projects for one body of work, such as a quarterly report with its reference files and brief.
• Memory as a convenience I do not rely on. I restate the important assumptions in each prompt.

The reason for that last point is practical. When an old assumption quietly leaks into a new answer, it is hard to spot. If you want the longer version, the ChatGPT memory deep dive covers how the feature behaves. And if you are choosing between assistants for the same prompt, the GPT-5 vs Claude comparison is a useful side-by-side, though test your own prompts rather than trusting any leaderboard.

How do I stop ChatGPT making things up?

A hallucination is a fluent answer the model has no basis for. Better prompts reduce it but never eliminate it, so build a checking habit.

1. Paste the source and say: “Use only this text. If the answer is not there, write ‘not in source’.”
2. After the answer, find each number, date, legal citation and quote in the original.
3. End the prompt with “mark the claims you are least sure about.”
4. For decisions that matter, ask the question a second way and see whether the answers diverge.
5. Open every link and look up every paper title. Invented references are a known failure.

For tax, legal or medical questions, treat the output as a draft and confirm with the official agency or a licensed professional.

What went wrong when one reader pasted everything in?

A composite example, based on a pattern I see often: a project coordinator, call her Dana, pasted a supplier quote and the company’s internal rate sheet into a chat and asked for a comparison table. The table looked great. It also had two problems. For rows where the rate sheet was silent, ChatGPT filled in a plausible “market rate,” and nothing in the table separated invented cells from real ones. And her IT team later pointed out that the rate sheet was marked as not to be shared with outside tools.

The fix had three parts. She replaced supplier names with Vendor A, B and C and converted prices to percentages of a baseline. She added “any value not present in the data must read ‘missing’; never estimate” to the constraints. For the most sensitive files, she moved to a model running on her own machine. If you want to automate that kind of routine step, the n8n AI agent automation guide is where I would start.

How do I keep improving my prompts?

Keep a notes folder called Templates. Once a week, skim the prompts you used and rescue the two or three that worked, replacing specifics with brackets. Within a month you will have a personal library, and that library is your real prompt-writing skill. The same discipline applies if you use AI for investing research: ask it for the questions to check in a company’s filings, not for a buy recommendation, and read our AI stocks investment guide with the same skepticism.

This article is general information. Features, settings and availability of ChatGPT change often, so confirm details in OpenAI’s Help Center, and follow your organization’s data policy before pasting work material into any online tool.