Four moves, not a hundred prompts. OpenAI’s own prompting guidance, translated out of developer language, comes down to this: structure the ask, show one example, pick the right answer mode, and correct instead of restarting. You can adopt all four in about ten minutes, and one of them contradicts the single most-shared piece of ChatGPT advice on the internet.
Most guides hand you a library of 100 copy-paste prompts. That is the wrong shape of help. Nobody memorizes 100 prompts, and the moment your situation drifts from the template you are back to staring at the empty box. What follows is one template and four habits that work on any request you will ever type.
Start here: the one template you actually need
OpenAI’s prompt engineering guide recommends four sections, in this order: Identity (who the model is and how it should sound), Instructions (the rules it has to follow), Examples (what a good answer looks like), and Context (the actual material). The order is doing most of the work there. Copy this and fill in the brackets:
You are [role, e.g. a plain-English editor for a small business owner].
Task: [what you want, in one sentence]
Rules:
- Length: [e.g. under 150 words]
- Format: [e.g. a short email / 5 bullets / a table]
- Tone: [e.g. warm but direct, no corporate filler]
- Avoid: [e.g. jargon, exclamation marks, made-up numbers]
Good example of what I want:
[paste one short sample, or one you liked from somewhere else]
Context:
[paste the email you're replying to, the notes, the document]Here is the same request before and after.
Before: write me an email to my client about the delay
After:
You are an account manager writing to a long-term client I like.
Task: tell them our delivery slips from Friday to the following Wednesday.
Rules:
- Length: under 120 words
- Format: email, no subject line
- Tone: straightforward, apologize once and move on
- Avoid: excuses, passive voice, "unfortunately"
Context: they've been a client for 3 years, the delay is our supplier's,
and we're covering the express shipping cost.Writing the second one takes under a minute, and it usually lands first time. Writing the first takes five seconds, and then you spend the next ten minutes fixing what came back.
Move 1: Say who, what, and what “good” looks like
The failure mode is named plainly in OpenAI’s own ChatGPT help pages: prompts need to be clear, specific, and carry enough context for the model to understand what you are asking, with ambiguity removed. Three things get left out, roughly in order of how much damage they do:
- Who it is writing for. “For my client” and “for my boss” and “for a stranger on the internet” produce three different emails. The model cannot guess.
- What format you want back. The same guidance asks you to be specific about outcome, length, format, and style. If you do not say “five bullets”, you get four paragraphs.
- What to avoid. Prohibitions do more work than requests. “No jargon” and “don’t invent statistics” each remove a whole category of bad answer.
The role line at the top is not decoration. Framing the model as a specific professional type, and describing how it should sound and what it is aiming at, is the first thing the prompt engineering guide asks for, because it narrows the vocabulary and the assumptions before a word gets written.
Move 2: Show one example instead of describing the style
Almost nobody bothers with this one. It is the fastest fix on the list. Describing a tone takes three adjectives and still misses; pasting one paragraph you like takes two seconds and hits.
OpenAI calls this few-shot prompting: include a handful of input and output examples showing the range of what you want. For an everyday request, one is plenty. Paste an old email you were happy with. Paste a competitor’s product description whose rhythm you like. Or dig out your own last report and say: match this format.
One caveat, and it runs opposite to what most guides tell you. OpenAI’s slower, more careful models are called reasoning models; in ChatGPT’s picker, that setting is Thinking mode, which Move 3 covers below. On those, OpenAI advises trying it with no example at all first (zero-shot) and adding one only if needed. So start plain. Add the example when the shape comes back wrong, not before.
Move 3: Pick the right answer mode
This single dropdown is worth more than any prompt phrasing, and most people never touch it. ChatGPT’s model picker is not a list of trivia, it is a speed-versus-depth dial. The version numbers change every few months; the modes do not.
| Mode | Use it for | What it costs you |
|---|---|---|
| Instant (fast) | Everyday questions, drafting, rewriting, translating, quick how-tos | Seconds |
| Thinking / higher reasoning effort | Long documents, math and logic, questions about files you uploaded, planning and decisions | Tens of seconds to minutes |
| Pro | The hardest tasks and longer-running work, where accuracy beats speed (higher plans only) | Minutes |
| Deep research | A written report drawn from many websites, with citations you can click | 5 to 30 minutes |
OpenAI frames the underlying split as planner versus workhorse: reasoning models handle ambiguity, accuracy, and planning, while faster models handle speed, cost, and well-defined execution. Most real work uses both. Draft in the fast mode, then paste the draft into a Thinking mode and ask what is wrong with it.
Deep research is the one people underuse. It works through many sources and returns a report with citations you can verify, which makes it the right tool for “should we switch suppliers” and the wrong tool for “rewrite this paragraph”.
Move 4: Correct it, do not restart it
When the answer comes back wrong, most people delete everything and retype a longer prompt. The docs say the opposite: refine iteratively. Start with a prompt, read the response, then adjust the wording, add context, or simplify, and go again.
Three follow-ups cover almost everything:
Too long. Half the length, keep the second point.Wrong tone. Match this instead: [paste sample]You changed the meaning of paragraph 2. Restore it and only fix the grammar.
Notice what those have in common: one specific change each. “Make it better” gives the model nothing to act on and it will simply reshuffle the words.
The second half of this move is decomposition, and the guidance here is to break a complex task into smaller focused prompts. So do not ask for a full business plan. Ask for the summary, approve it, then ask for the market section using the approved summary. You catch a wrong turn on page one instead of page nine.
Set it up once so you stop repeating yourself
Everything above is per-chat. These three features make it permanent, and they are where the biggest lasting gain sits for someone who uses ChatGPT most days.
| Feature | What it does | When to use it |
|---|---|---|
| Custom Instructions | A standing note applied to every new chat: what ChatGPT should know about you, and how you want it to answer | Once, now. Available on all plans |
| Memory | Carries facts across conversations, both memories you ask it to save and insights drawn from your chat history | Leave on; say “remember that…” for things you repeat |
| Projects | A workspace holding related chats, uploaded files, and its own separate instructions, with built-in memory | Any effort lasting more than one conversation |
OpenAI documents Custom Instructions at up to 1,500 characters on Free and Go plans and 5,000 on Plus, Pro, Business, Enterprise, and Edu. Here is a starter worth pasting into it today:
About me: I run a 6-person design studio in Warsaw. I write to clients,
suppliers, and my team. English is my second language.
How to answer:
- Plain language. No corporate filler, no "delve", no "in today's landscape".
- Lead with the answer, then the reasoning.
- If my request is ambiguous, ask one clarifying question before answering.
- If you're not confident about a fact, say so instead of guessing.
- Match my level: explain unfamiliar terms in one line the first time.Rewrite those five lines in your own situation and every chat you start from now on begins better than the one before it.
Stop copying these four tips from the internet
The most-shared ChatGPT advice was written for models that no longer exist. Some of it now actively costs you quality.
| Popular tip | What OpenAI’s docs actually say | Do this instead |
|---|---|---|
| ”Add ‘think step by step‘“ | Instructions that ask for step-by-step thinking (chain-of-thought) are listed among the things to avoid on Thinking-mode models; keep prompts simple and direct | Switch to Thinking mode. Do not ask for thinking, select it |
| ”Ask it to explain its reasoning” | Also listed among the techniques to avoid on Thinking-mode models | Ask for the sources, or ask what would make the answer wrong |
| ”Stack personas and pile on examples” | Try zero-shot first, then few-shot if needed. Keep prompts simple and direct | Start plain. Add one example only when the shape comes back wrong |
| ”Tell it to self-grade until it scores 9/10” | Not in any OpenAI guidance. The documented move is giving explicit constraints and specific success criteria up front | Write the success criteria into the prompt: “under 120 words, no jargon, one apology” |
That last row is the pattern behind all four. Community tricks that ask the model to fix a prompt after the fact are usually working around a constraint you could have stated in the first place. One popular tactic does survive contact with the docs, though its usual form is overwrought: asking ChatGPT to ask you clarifying questions before it answers. It genuinely helps on a fuzzy request. It is also a sign your prompt did not state its constraints, which is cheaper to fix at the source.
Why your last answer was bad
| Symptom | Cause | Fix |
|---|---|---|
| Generic, could be about anything | No role, no audience, no context | Add the Identity line and paste the source material |
| Right facts, wrong shape | Format never specified | State length, format, and structure explicitly |
| Tone is off and stays off after two tries | You described the tone instead of showing it | Paste one sample and say “match this” |
| It ignored half of the long document you pasted | Models use the start and end of a long input far more reliably than the middle, well before any size limit | Paste only the relevant sections, and put your instructions at the top and repeat the key one at the bottom. Background: lost in the middle |
| Confident sentences containing invented facts | It was asked for an answer, not for sources | Ask in a mode that cites sources, or add “if you are not sure, say so” |
| It drifted after five good messages | The thread accumulated too much contradictory instruction | Start a fresh chat and paste in only the approved output |
Check it before you trust it
ChatGPT will state something wrong in the same confident register it uses for something right. That does not make it unreliable for work, it makes it a first draft with an unreliable footnote section.
The practical rule: verify anything that carries consequence, and let the tool help you do it. Ask for sources and click them. Use Deep research, which returns citations for its claims, when the answer will inform a decision rather than a sentence. And treat anything about your own numbers, contracts, health, or legal position as a draft for a human to check, never an answer.
There is a second check worth running before you hit send rather than after. Whatever you paste leaves your machine. Client contracts, other people’s personal details, anything under an NDA: find out what your organization allows before that text goes into the box.
Do this in the next ten minutes
- Open Settings and write five lines of Custom Instructions. Two minutes.
- Copy the template above into a note you can reach quickly. One minute.
- Take the last mediocre answer you got, and instead of retyping, send one specific correction. Two minutes.
- Take your next real task and pick the mode before you type the prompt. Ongoing.
That is the whole method. If you want to see where the same discipline goes when you push it further, the underlying idea is that what you put in the window matters more than how big the window is: context engineering beats a bigger context window.