Prompts that work

Prompts that work

A field guide for anyone new to talking to AI: nine skills, each shown with a bad prompt, the real answer it got, and a better prompt you can copy. Plus good habits, traps, and how the advice has changed.

For first-time prompt writersSources checked 22 Oct 2026About 25 minutes

A friendly robot at a desk: on the left a scribbled, tangled note; on the right a human hand holds out a neat, glowing page with a check mark. A trail of sparkles links the two.

One prompt, taken apart

Here is how most of us start, and then the same request with six skills added. It is longer, but every line removes something the AI would otherwise have to guess.

Vague
"Write an email to my teacher about my homework."

Which homework? What happened? What do you want from the teacher? How formal should it be? The AI has to guess all of it, so you get a polite, empty template.

Clear
Write an email to my chemistry teacher asking for two more days on my lab report.Clarity I was off sick on Monday and Tuesday and missed the experiment.Context She is strict but fair, so keep it polite and honest, not grovelling.Tone My notes are between the tags: <notes> … </notes>Structure Under 100 words, with a subject line, and suggest Friday as the new date.Format Before you finish, check that every fact in my notes is in the email.Check
  • Clarity: one job, said with a verb.
  • Context: what happened, what you want.
  • Tone: who reads it, how it sounds.
  • Structure: your notes, fenced off.
  • Format: length and shape you can count.
  • Check: one concrete thing to verify.

Nine skills, thirty rules

Each skill starts with the same three steps: a bad prompt, the answer it really got, and a better prompt. The answers come from DeepSeek V4.1 Flash, run on 22 October 2026 and shortened where you see "…". Then come the skill's rules, each with the reason it works and a badge showing where the advice comes from.

  • Research measured in a published study
  • Official guide advice from OpenAI, Anthropic or Google
  • Field practice common habit, not formally tested

Clarity

Say exactly what you want. The AI fills every gap with the most average guess, so leave fewer gaps.

The robot holds a magnifying glass over a fuzzy grey cloud; inside the lens it becomes a crisp coral star.

Try it

  1. Bad prompt

    "Help me with my history essay."

  2. What it got

    Absolutely—I can help you plan, outline, draft, or revise a history essay. To give you useful help, tell me:

    1. Essay prompt/question (exact wording if possible)
    2. Topic/time period/region
    3. Length and level …

    … A strong history essay usually needs:

    • A clear argumentative thesis
    • Context for the time/place …

    Seven questions back and some general tips. No help yet, because there was nothing to work on.

  3. Better prompt

    "List the 3 biggest weaknesses of my history essay thesis below, one line each. Then suggest a stronger thesis of at most 30 words for a 1,000-word essay. Thesis: 'The French Revolution happened because people were angry.'"

The rules

Start with one clear job and an action verb

"List", "rewrite", "compare" and "explain" tell the AI what to produce. "Help me with" makes it guess, or send you questions instead.

Official guideAnthropic, 2025

Swap vague words for things you can count

"Short", "good" and "detailed" mean something different to everyone. "3 points", "under 30 words" or "5 ideas" can be checked. A 2026 study found vague prompts give answers that vary more.

ResearchPecher et al., 2026

Pass the new-colleague test

Picture handing your prompt to someone who just joined your class. If they would need to ask "for what?" or "who is it for?", the AI needs to know too. Say what you want, who it is for and why.

Official guideOpenAI Academy, 2026

Ask for extra effort when you want it

Today's models do what you ask, quite literally. If you want more ideas, more detail or a deeper answer, say so.

Official guideAnthropic, 2025

Context

Give it the facts only you know. The AI has never met your mum, your teacher or your bug.

The robot watches a hand place the last piece of a jigsaw that shows an arrow in a bullseye.

Try it

  1. Bad prompt

    "What should I get my mum for her birthday?"

  2. What it got

    A few good directions depending on what she's like:

    • Pampering: spa voucher, nice hand cream/body care set, silk scarf, or a cozy robe.
    • Experience: afternoon tea, theatre tickets, a cooking class …

    … If you tell me her age, interests, and budget, I can narrow it down to 5 specific ideas.

    Ideas for any mum. The AI even asks for the facts it is missing.

  3. Better prompt

    "Suggest 3 birthday presents for my mum. She is turning 45, loves gardening and true-crime podcasts, and already owns every gardening tool. My budget is 25 euros, and it has to arrive by Saturday or be something I can make myself."

The rules

Say why you are asking

The same question needs a different answer for a test tomorrow, a school project or plain curiosity. One line about your goal changes the depth and the angle.

Official guideAnthropic, 2025

Give it the facts it can't know

Your assignment brief, your error message, your teacher's comments: none of it is in the AI's training. Paste it in, then you can check the answer against it.

Official guideOpenAI Academy, 2026

For questions about a document, ask for quotes first

Ask it to copy the sentences that answer your question before it explains. If it can't find a quote, the answer probably isn't in the text.

Field practice

Structure

Keep your instructions and your material apart, so the AI never mixes them up.

The robot files loose papers into a cabinet of colour-coded drawers.

Try it

  1. Bad prompt

    "correct this: i think school should start at 9 because teens need more sleep and we learn better. what do you think?"

  2. What it got

    I think school should start at 9:00 because teens need more sleep and we learn better. What do you think?

    It corrected the question as well, and never answered it. It couldn't tell where your text ended.

  3. Better prompt

    "Correct the spelling and grammar of the text between the <text> tags. Keep my words and my opinion. <text>i think school should start at 9 …</text> Then, separately, tell me in two sentences how convincing the argument is."

The rules

Fence off anything you paste

Put pasted text inside tags like <text> … </text>, and say what to do with it. Anthropic recommends marking pasted text this way, so the AI is less likely to treat it as instructions.

Official guideAnthropic, reported by SEJ, 2026

Give each part a name you can point to

With a heading like "My essay:" or a tag like <feedback>, you can say "use only the feedback". Anthropic says clear headings work as well as tags for most prompts.

Official guideAnthropic, 2025

Long material first, your question last

With a long text, put the text first and your instructions after it, so the last thing the AI reads is what you want.

Field practicePhilipp Schmid, 2025

Use short labelled sections for long prompts

For a big task, write a few labelled parts: goal, context, format, limits. OpenAI's own guide uses these four blocks, and says short prompts often need none of them.

Official guideOpenAI, reported by The Decoder, 2026

Examples

Show, don't describe. One good example teaches the tone and the format faster than a paragraph of rules.

The robot holds up a card it drew to match three sample cards pinned on a board.

Try it

  1. Bad prompt

    "Write 5 Instagram captions for my band's first gig."

  2. What it got

    1. First gig. First nerves. First chance to play our songs for you. See you at [venue] on [date] 🎸🔥 #FirstGig #LiveMusic #[BandName]

    2. We've been practicing for this moment, and now it's real: OUR FIRST SHOW. … 🖤🎶 #BandLife #SupportLocalMusic …

    Could be any band. Blanks to fill in, emoji and hashtags, in nobody's voice.

  3. Better prompt

    "Write 5 Instagram captions for my band's first gig (Friday, The Cellar, 8 pm). Match the style of two we posted before: 'new song. old amp. same bad haircuts. friday.' and 'we learned 3 chords and got carried away'. Lowercase, under 12 words, no hashtags."

The rules

Show one to five examples of a good answer

Google and Anthropic both recommend examples, and Anthropic says start with one. A 2026 study found they help some models and hurt others, so start small and compare.

Official guideGoogle, 2026ResearchDwivedi & Soni, 2026

Make examples varied, and true to your rules

The AI copies everything in an example: its length, its tone, even its mistakes. Vary the topics, keep the format the same, and make sure each one follows your own rules.

Official guideGoogle, 2026

Role & audience

Say who the answer is for and how it should sound. One line sets the words, the depth and the tone.

The robot in a graduation cap teaches a child at a small chalkboard, with a chef's hat and a detective's hat floating nearby.

Try it

  1. Bad prompt

    "Explain how vaccines work."

  2. What it got

    Vaccines work by teaching the immune system to recognize and fight a specific pathogen … A vaccine contains an antigen … Immune cells recognize the antigen as foreign. They begin an immune response, including activating B cells and T cells …

    Correct, and about 300 words long. Fine for a biology test, useless for calming a scared 9-year-old.

  3. Better prompt

    "My little brother is 9 and scared of his vaccine tomorrow. Explain how vaccines work in 4 short sentences he would understand, with one comparison to something from everyday life. Keep it calm and reassuring."

The rules

Name the reader

"For a 9-year-old", "for my teacher" or "for someone who knows the basics" tells the AI which words to use, how deep to go and what to skip.

Official guideOpenAI Academy, 2026

A role sets the voice, not the expertise

"You are a world-class expert" adds no knowledge. Two studies from 2025–26 found expert roles rarely make answers more accurate, and sometimes less. A role can still help with tone.

ResearchBasil et al., 2025Hu et al., 2026

Put rules that never change in your settings

If you always want British spelling or short answers, put it in your chat app's custom instructions instead of typing it every time. OpenAI's guide says the same.

Official guideOpenAI, reported by The Decoder, 2026

Format

Describe the finished answer: its shape, its length, what goes in it. Then it arrives ready to use.

The robot presses a star-shaped cutter into dough next to a tray of identical star cookies.

Try it

  1. Bad prompt

    "Make me a revision plan for my exams."

  2. What it got

    I can make you a proper revision plan, but I need a few details first:

    1. What exams/subjects do you have?
    2. What are the exam dates? …
    3. Do you prefer short daily sessions or longer blocks?

    In the meantime, here's a flexible revision framework …

    | Mon | Weak subject: active recall | Strong subject: practice questions | …

    Eight questions and a blank template. You still have to make the plan yourself.

  3. Better prompt

    "Make a revision plan as a table with the columns Day, Subject, Topic and Minutes. Exams: maths 12 June, biology 14 June, history 17 June. I can do 90 minutes on weekdays, 3 hours at weekends. Weakest topics: quadratics and the cell cycle. Start Monday 1 June. Reply with only the table."

The rules

Describe the shape of the answer

A table with named columns, five bullet points, an email with a subject line: say what the finished answer looks like, and you won't have to reformat it.

Official guideOpenAI, reported by The Decoder, 2026

Ask for length in units you can count

"Short" is a guess. "One sentence", "under 100 words" or "4 bullet points" is a target you can check. Google's own guide uses "Summarize this text in one sentence".

Official guideGoogle, 2026

Say what to do, not only what to avoid

"Write in short paragraphs, because I read on my phone" works better than "NEVER use bullet points": it gives a target and a reason. Keep "never" for real limits.

Official guideAnthropic, 2025

Write your prompt in the style you want back

The way you write rubs off on the answer. A prompt full of bullet points tends to get bullet points back, and a chatty prompt gets a chatty answer.

Official guideAnthropic, 2025

Reasoning

Help the AI think in a way you can follow and check, whatever model you are using.

The robot climbs a staircase of colourful blocks towards a glowing light bulb.

Try it

  1. Bad prompt

    "Just give me the answer: 2x² − 8x + 6 = 0"

  2. What it got

    x = 1 or x = 3

    Right, but nothing to check or learn from. You are stuck on the next one.

  3. Better prompt

    "Solve 2x² − 8x + 6 = 0 step by step, with one line explaining each step, so I can follow it. Then check both answers by putting them back into the equation."

The rules

Everyday chat models: ask for the working

If your app has no "thinking" mode, asking for the steps helps on tricky problems, and lets you see where it went wrong.

Official guideAnthropic, 2025

Reasoning models: give the goal, not the steps

Models that think before answering plan better on their own. Describe the result you want, and spell out a process only when the process itself matters.

Official guideOpenAI, reported by The Decoder, 2026

Iteration

Your first prompt is a draft. Read the answer, find what's missing, and say exactly what to change.

The robot polishes a dull rock inside a loop of arrows, and the rock becomes a shining gem.

Try it (after a first draft of a class-rep speech)

  1. Bad prompt

    "Try again."

  2. What it got

    Here's a different version—more energetic and concrete: … If you choose me, I'll focus on three things: keeping everyone informed so no one misses important updates, making sure your concerns actually get raised—not just heard … I can promise I'll be approachable, reliable, and honest …

    It says "more concrete", but the promises are just as vague. "Try again" never said what was wrong.

  3. Better prompt

    "Good start. Keep the first line, but swap the general promises for two concrete ones: a shared revision folder and a class vote every month. Leave everything else as it is."

The rules

Treat a bad answer as feedback

A weak answer shows you what your prompt left out. Add that, in the same chat, instead of starting over.

Official guideOpenAI Academy, 2026

Change one thing at a time

If you change three things and the answer improves, you don't know which one helped. OpenAI's guide to switching models gives the same rule: "Make one change at a time."

Official guideOpenAI, 2025

Diagnose before you shout

When the AI ignores something, look for a missing fact or two rules that clash before you reach for CAPITALS. OpenAI warns that clashing rules can unsettle its newest model more than missing detail.

Official guideOpenAI, reported by Decrypt, 2026

Save prompts that work, and retest on new models

Keep a note of prompts that worked well. When your app switches to a new model, try them again: Anthropic's guide for its newest model asks developers to re-check setups carried over from older ones.

Official guideAnthropic, reported by SEJ, 2026

Agents & coding

When the AI can act (edit files, run code, browse the web), your prompt becomes a work order.

The robot types on a laptop beside a toolbox, a ticked checklist, gears and a green check-mark badge.

Try it (with a 12-line guessing game pasted below the prompt)

  1. Bad prompt

    "my python game doesnt work. fix it"

  2. What it got

    Your code runs, but it doesn't give feedback, and it crashes if the player types something that isn't a number. Here's a fixed version: [the whole game, rewritten] … What I fixed/improved:

    • Prevents crashing on non-number input.
    • Checks that the guess is between 1 and 3.
    • Prints "Correct!" or "Wrong!" feedback …

    It guessed what "doesn't work" meant and changed five things. On a real project, that is five changes to review.

  3. Better prompt

    "Fix one thing in my Python game: after a wrong guess, the player never learns the secret number. Done means it prints 'Wrong! It was N.' after every wrong guess. Change nothing else, then tell me how to test it."

The rules

Define the scope and what "done" means

Say what to change, what to leave alone, and how you will know it is finished. OpenAI's latest guide tells you to keep exactly that: the result, the success test and the hard limits.

Official guideOpenAI, reported by Decrypt, 2026

Give it a check it can run itself

A test, a build or a screenshot to compare. Anthropic calls verification its single most useful Claude Code tip. Without it, you become the one who catches every mistake.

Official guideAnthropic, 2026

Write the house rules down once

Coding agents read a file called AGENTS.md in your project: how to run the tests, which style to follow. It is now a shared open standard, and OpenAI's Codex is trained to follow it closely.

Official guideLinux Foundation, 2025OpenAI, 2026

What the measurements say

Most prompting advice comes from experience. These three results come from studies published in 2026, some on older models. They are small tests, so treat them as a direction, not a law.

Examples helped one model and hurt another

Sorting 200 news articles into 4 topics. Score out of 1 (macro-F1).

  • No examples
  • Two examples

Two examples was the small model's best setting and the large model's worst. Try your prompt with and without examples.

Dwivedi & Soni, 2026, Table II · arXiv 2607.22969

A little context changed what AI agents bought

AI shopping agents choosing a fitness tracker: how much more often they picked one product, 200 runs per setup.

Every bit of context counts. Give the AI what matters and leave out the rest.

Kumar et al., 2026, Prompting Science Report 6 · Wharton

Shorter instructions did better

OpenAI's coding-agent tests: lean prompts versus long ones packed with repeated rules.

Bars show the reported range. Longer prompts are not automatically better.

OpenAI, reported by Decrypt, July 2026 · Decrypt

Good habits, and traps to avoid

The nine skills are about one prompt. These habits are about using AI well every day.

Do

  • Brief it like a teammateSay the goal, what a good result looks like and how you'll check it. Ethan Mollick calls clear goals the core of working with AI.
  • Check facts, quotes and sourcesIn a 2026 survey of 63 students, made-up citations and false facts were the main mistakes they reported. Open every source it gives you.
  • Use it to learn, not to skip learningAsk for the steps, a hint or a quiz on your own notes. Your teacher marks what you understand.
  • Keep a prompt notebookSave the prompts that worked, and reuse them.
  • Keep agents on "ask me first"Before an AI sends, buys or deletes anything, it should ask you. Hidden text on a web page can try to trick it.

Avoid

  • Sharing private detailsLeave out passwords, addresses and other people's secrets. A chat is not a diary.
  • Rules that clash"Be very detailed" plus "keep it short" makes the answer wobble. Pick one.
  • Pasting everything you haveIn a 2026 study, each extra layer of reviews and remembered details made AI shopping agents less predictable. Give it what matters.
  • "Try again" with no reasonSay what was wrong, or you'll get the same mistake in new words.
  • Trusting the confident toneAn AI sounds just as sure when it's wrong. Fluent is not the same as true.
Four stations in a circle (a note, a magnifying glass, a light bulb and a paper plane) around the robot giving a thumbs up.

The practice loop

  1. Write your prompt the way you normally would.
  2. Check it against the nine skills.
  3. Fix the two or three biggest gaps, then send it.
  4. Read the answer as feedback and loop back if needed.

How the advice has changed

Many famous tricks were patches for weaker models. Today's models follow instructions more closely and reason on their own, so some old patches no longer help, and some get in the way.

Then"Add 'think step by step' to every prompt."

Less needed

NowMany apps now have a thinking mode that reasons for you, though free plans may not. Ask for the steps when you want to see them, or when there is no thinking mode.

Anthropic, 2025

Then"WRITE IMPORTANT RULES IN CAPITALS."

Retired

NowOlder models skipped instructions, so shouting helped. Newer Claude models may overdo it instead, using a tool too often when told it is CRITICAL. Anthropic now advises "Use this tool when…" instead of "CRITICAL: You MUST…".

Anthropic, quoted by Corti, 2026

Then"Start with 'You are a world-class expert'."

Mostly myth

NowTwo studies found accuracy rarely improves, and sometimes drops. A role can still shape the tone, and Anthropic calls heavy role prompting "often unnecessary" now.

Basil et al., 2025Hu et al., 2026Anthropic, 2025

Then"A longer, more detailed prompt is always better."

Reversed

NowIn OpenAI's coding-agent tests, leaner prompts with fewer repeated rules and pointless examples scored about 10–15% higher. A short prompt often works.

OpenAI, reported by Decrypt, 2026

Then"Always tell it to double-check everything."

Be specific

NowAnthropic says Claude Opus 5 checks its own work, and leftover "verify" lines make it over-check. Ask for one concrete check instead.

Anthropic, quoted by Corti, 2026

Then"Wrap every part of your prompt in XML tags."

Optional now

NowAnthropic says modern models understand structure without them: clear headings and plain words work as well for most prompts. Tags still help to fence off pasted text.

Anthropic, 2025

The pattern: tricks fade with each new model. What lasts is a clear goal, the facts, the shape of the answer, and a way to check it.

Before you press Enter

Ten quick checks, taken from the thirty rules. Tick them off on your next prompt; nothing is saved.

0 of 10 checked

What to remember

A prompt is a brief for a fast, literal helper that knows nothing about your situation.

All nine skills close the same gap between your head and the screen. Name the job with a verb. Add the facts only you know. Say who it is for and what the answer should look like. Then read the reply as feedback, and change one thing at a time.

Tricks come and go with every new model; a clear goal and a way to check the answer don't. And the answer is still yours: check the facts, keep private things private, and use AI to learn, not to skip learning.

Sources

Every source is dated between October 2025 and October 2026. Pages without a date were left out.

Official guides

Research

Explainers and reporting

How this was checked. Each page was opened on 22 October 2026 and its date read on the page itself. The "What it got" answers are real replies from DeepSeek V4.1 Flash, shortened. The charts are drawn from the studies' own numbers. The illustrations were made locally with MOSS-Image.