Get Started
Use AI properly

Mission 2 of 11

Ten prompts that turn a bland chatbot into a genuinely useful one.

Same chatbot, better brief, completely different answer. Try each weak prompt against its strong version, then keep the kit to paste into ChatGPT, Claude or Gemini.

45 minBeginnerFreeFree mission

What you'll do

Compare weak and strong prompts

Send each pair yourself and see exactly what changes in the answer, and why.

Fix any weak prompt

Use the six part anatomy to fix a prompt in under a minute, and the what went wrong map to pick the right one.

Build your own reusable prompt

Turn a job you really do every week into a prompt you only write once.

  • Ten copy and paste prompts that work in any chatbot
  • A kit you can reuse for the rest of your life
Before you start
  • Any chatbot open in a browser tab or on your phone: ChatGPT ↗, Claude ↗ or Gemini ↗
  • About 45 minutes. You'll send most of these for yourself, because seeing the difference is what makes it stick
  • A real task you've been putting off. An email, a decision, a plan. You'll use it in the exercises
Your progress0/6
● What you'll have in 45 minutes
Preview
An illustration of the whole course in 13 seconds: the same request, asked two ways.
01/06Prompt 1: Give it a job and the context0 done
STEP 01 of 06~9 min

Prompt 1: Give it a job and the context

Get your first properly useful answer in the next few minutes, and see why the last ones weren't.

Most people use a chatbot like a search box, get something bland back and decide AI is overrated. Here you'll see why, and fix it.

Set up and send the weak version

  1. Open your chatbot. Pick ChatGPT, Claude or Gemini in a new tab. Everything in this course gets pasted into the main message box.

    The real Gemini home page, signed out. The box marked 1 is where your prompt goes. Other chatbots put theirs in the same sort of place.

    Prompt 1 is for any time the answer has to fit your situationYou'll start with a deliberately thin version, so you can see what goes wrong.

  2. Paste this weak version and send it.

    Weak version
    Write an email to my landlord about the mould in my flat.

    You'll get something like this"Dear Landlord, I hope this email finds you well. I am writing to inform you of a mould issue in my flat. Please could you look into this when convenient."

Why the answer was bland

Look at what it had to guessIt wasn't lazy. It had one fact to work with, so it wrote the most likely email from the most average tenant.

A chatbot predicts what should come next. Give it nothing specific and the likeliest answer is the average one.

Send the strong version

  1. Open a new chat and send the strong version. A new chat stops the first attempt colouring the second.

    Strong version: paste into any chatbot
    You're helping a tenant in England write to their landlord.
    
    Situation: I rent a one-bed flat in Leeds. Black mould appeared on the bedroom wall three weeks ago. I rang the letting agent on 6 October and nothing has happened. My daughter is 4 and has started coughing at night.
    
    Task: write the email to my landlord. The goal is an inspection booked within 7 days.
    
    Rules: polite but firm, no threats, under 150 words, include a subject line. Don't invent legal claims. Use [square brackets] for anything you don't know.

    You'll get something like thisa subject line naming the problem, then the mould, the date you rang, your daughter's cough and a clear request for an inspection within 7 days. Your wording will differ, but it should read like a tenant who means it.

    Tip
    Why it wins. Every fact you add removes a guess. The job points it at the right kind of writer, the goal tells it what the email is for, and the rules stop it padding or bluffing. The square brackets are the sneaky part: they give it a way to admit a gap instead of making up your address.

The six parts of a prompt

Find each of the six parts in the strong version you just sent.

Prompt anatomy. You won't need all six every time, but when an answer disappoints, one of these is usually missing.

See the same advice from the sourceClaude's makers tell people to treat it like a very smart new starter, and the same habit works in any chatbot. Read the two highlighted parts.

The real Prompting best practices ↗ page in Claude's docs.
Tip
The new starter test. Before you send a prompt, imagine handing it to someone clever on their first day. If they'd come back with five questions, put the answers in the prompt.

Your turn

  1. Take the last thing you asked a chatbot, or the task you've been putting off, and rewrite it with at least four of the six parts. Send it in a new chat.

    Tip
    Stuck? Start with context. Write down what you'd have to explain to a friend before they could help you, then put a task sentence underneath.
Checkpoint
You've got one answer you could actually use, and you can name the parts of a prompt without looking.
If it breaksMy two answers look nearly the same + 2 more
My two answers look nearly the same
Check you opened a new chat for the second one. If you did, your weak prompt probably already contained plenty of detail, which is good news. Try the pair on a task where you left more in your head.
It wrote far more than I asked for
Reply with the limit again, for example Cut this to 100 words and keep the request. You don't need to start over.
The chatbot wants me to sign in
Some chatbots need a free account, and what you can do signed out varies by product and country. If you'd rather not sign up today, try one of the others.
Still stuck? Ask the community
STEP 02 of 06~8 min

Prompts 2 and 3: Get what's in your head into the chat

Fix the two hardest cases: when you can't say what you want, and when you can't describe how it should sound.

Prompt 1 gave the chatbot facts you already had. These two cover the harder cases.

Prompt 2: Make it interview you

Use this when you want something tailored but don't know what to tell it.

  1. Start a new chat and send the weak version.

    Weak version
    Plan me a 3 day trip to Lisbon.

    You'll get something like thisDay 1 is the Alfama and the castle, day 2 is Belém, day 3 is a trip to Sintra. A decent plan for nobody in particular.

  2. Open a new chat and send the strong version. Then answer its questions until it writes the plan.

    Strong version: paste into any chatbot
    I want a 3 day trip to Lisbon. Before you plan anything, interview me. Ask one question at a time, wait for my answer, and stop when you know enough to write a plan only I would want. Start with the question that would change the plan the most.
    Demo
    An illustration of how the conversation goes. One question at a time, and each answer changes the next question.
    Tip
    Why it wins. You don't have to work out what matters, because it asks. "One at a time" keeps the questions sharp and lets each one build on your last answer. "The question that would change the plan the most" stops it opening with trivia.

Prompt 3: Show an example

Use this when the voice or layout matters and adjectives aren't working.

  1. Send the weak version in a new chat.

    Weak version
    Write an Instagram caption for my bakery's new cinnamon buns.

    You'll get something like this"Indulge in our irresistible new cinnamon buns! Freshly baked with love and swirled to perfection." Then a row of emojis and hashtags. It could be any bakery on earth.

  2. Open a new chat and send the strong version. Read the result out loud next to the two examples.

    Strong version: paste into any chatbot
    Write an Instagram caption for my bakery's new cinnamon buns. They're out Saturday at 8am and we're only making 40.
    
    Match the voice of these two past captions exactly:
    
    "Sourdough's back. Yes, the burnt-edge one. Door opens at 8. Don't be late, Dave won't save you one."
    
    "We made too many brownies. That is a lie. There's no such thing. Come and prove us right."
    
    Output only the caption.

    You'll get something like this"Cinnamon buns. Saturday. 8am. We made 40, which sounds like a lot until you meet Dave." Short sentences, a dry joke, and the same running character.

    Tip
    Why it wins. "Funny and casual" means something different to everyone. Two examples carry the sentence length, the humour and the rhythm in one go, with no describing needed. Anthropic's guide suggests three to five examples for best results, so add more if the voice is drifting.

Your turn

  1. Find two things you've written that you liked, such as messages, posts or emails. Paste them into prompt 3 and ask for something new in that voice.

    Tip
    Stuck? Search your sent emails or your own social posts. Pick examples that are alike in the way you want copied. If one is formal and one is jokey, it will split the difference.
Checkpoint
You've had a chatbot ask you questions and write something based on your answers, and you've got one piece of writing that sounds like you.
If it breaksIt asked all its questions in one big list + 2 more
It asked all its questions in one big list
Reply: One at a time, please. Start with the most important. It will usually fall straight into line.
It keeps interviewing me and never finishes
Reply: That's enough. Write the plan now, and list any assumptions you made.
It copied my example too closely
Reply: Keep the voice but use completely new jokes and phrasing.
Still stuck? Ask the community
STEP 03 of 06~8 min

Prompts 4 and 5: Stop the guessing and the rambling

Get answers based on your documents, in a shape you can use straight away.

So far you've been feeding the chatbot your situation and your voice. Now you'll feed it a document, and control the shape of what comes back.

Prompt 4: Stick to the source

Use this whenever the answer lives in a specific document, such as a contract, policy or set of notes.

  1. Send the weak version in a new chat.

    Weak version
    How much notice do I need to give to cancel my gym membership? Is there a fee, and can I pause it instead?

    You'll get something like this"Notice periods vary, but many gyms require around 30 days. Check your contract." It has never seen your contract, so all it can give you is what's typical.

  2. Open a new chat and send the strong version. The contract text is made up for practice, so it's safe to paste.

    Strong version: paste into any chatbot
    Answer my questions using only the contract text below. Quote the exact line you used for each answer. If the text doesn't cover something, write "Not in the document" and don't guess.
    
    Questions:
    1. How much notice do I need to give to cancel?
    2. Is there a fee for leaving?
    3. Can I pause my membership instead?
    
    Contract text:
    "Either party may end this membership by giving two full calendar months' written notice. An administration fee of £25 applies if the membership is ended within the first 12 months. Fees are collected by direct debit on the 1st of each month."

    You'll get something like thistwo full calendar months, with the line quoted. A £25 fee if you leave within the first 12 months, with the line quoted. And for pausing: "Not in the document".

    Tip
    Why it wins. A chatbot is built to be helpful, and with no source that often means sounding sure. The source gives it something better than memory, the quotes let you check its work in seconds, and "Not in the document" gives it a way out. That third question is the proof: a good reply refuses to invent a pause policy.
    Warning
    On anything that matters, such as contracts, money, health or legal questions, find each quoted line in the original yourself. If it isn't there word for word, don't trust the answer.

Keep private documents private

  1. Open a private chat before you paste anything private.

    In Claude, open a new chat and click the ghost icon in the top right. In Gemini, sign in and click Temporary chat next to New chat. ChatGPT calls its version Temporary Chat too.

    Warning
    Private modes keep a chat out of your history, but it still goes to the company and is held for a while. Never paste passwords or card numbers into any chatbot, and swap names and account numbers for placeholders such as [TENANT] before you paste a document.

Prompt 5: Force a format

Use this when you'll act on the answer and don't want to dig through paragraphs.

  1. Send the weak version in a new chat.

    Weak version
    What are the pros and cons of renting alone versus a flatshare?

    You'll get something like thisa long list of pros and cons under two headings, an introduction you didn't need, and a closing line saying it depends on your priorities.

  2. Open a new chat and send the strong version. You should be able to read the answer in one glance.

    Strong version: paste into any chatbot
    Compare renting alone with a flatshare for someone in their first job.
    
    Format: one table with 3 columns (Factor, Renting alone, Flatshare) and exactly 5 rows: monthly cost, privacy, bills and admin, social life, flexibility to move. Maximum 8 words per cell.
    
    Under the table, write one sentence saying which you'd pick if saving money matters most. No introduction and no closing paragraph.

    You'll get something like thisa neat five row table, then one sentence recommending the flatshare if saving money comes first. No warm up and no sign off.

    Tip
    Why it wins. A format forces decisions. A limit of 8 words per cell makes it pick the point that matters, and asking for a recommendation stops it hiding behind "it depends".

Your turn

  1. Find a long terms page, an email thread or some notes that have nothing private in them. Use prompt 4 to ask three questions, and make sure one of them isn't answered in the text.

    Tip
    Stuck? A public returns policy from any online shop is perfect. For the trick question, ask something it won't cover, such as what happens if you're returning from another country.
Checkpoint
You've seen a chatbot quote your source and say "Not in the document", and you've turned a question into a table you can scan.
If it breaksIt answered the trick question anyway + 2 more
It answered the trick question anyway
Reply: Show me the exact line that says that. If it can't, tell it to answer again using only the text. Then re-read the prompt and check the "Not in the document" line is still there.
My table came out as messy text
Reply: Show that as a markdown table. If the table still won't display, ask for a plain list with one item per line.
My document is too long to paste
Paste the section that matters. Most chatbots also let you attach a file, so look for an attach or plus button near the message box.
Still stuck? Ask the community
STEP 04 of 06~7 min

Prompts 6 and 7: Make it think before it writes

Get better ideas, and catch wrong turns before they cost you a whole draft.

Until now the chatbot has produced one answer and stopped. These two prompts slow it down on purpose, so the thinking happens before the writing.

Prompt 6: Ask for options, then critique

Use this for names, headlines, ideas and decisions, anywhere the first answer is rarely the best.

  1. Send the weak version in a new chat.

    Weak version
    Give me a name for my dog walking business.

    You'll get something like thisa handful of pun names along the lines of Pawsome Walks and Happy Tails, with no view on which one is any good.

  2. Open a new chat and send the strong version. Watch how the work moves through four stages.

    Strong version: paste into any chatbot
    I run a dog walking business in Bristol, mostly for busy professionals.
    
    1. Give me 10 name options, each in a noticeably different style.
    2. Score each one out of 5 for: easy to say out loud, easy to remember, doesn't sound like every other dog walker.
    3. Cut to the top 3 and tell me the biggest weakness of each.
    4. Pick one and tell me why in two sentences.
    Four stages take you from a pile of ideas to one decision you can defend.
    Tip
    Why it wins. The first answer is the likeliest answer, which is why it feels familiar. Ten options push it past the obvious, "noticeably different" stops ten versions of one idea, and your scoring criteria make it judge rather than just generate.

Prompt 7: Plan before doing

Use this for anything big, such as reports, presentations, long emails or codeRead the weak version. You don't need to send this one, because a full deck takes a while.

Weak version
Write a 10 slide presentation on why our team should try a four day week.

You'd get something like thisten finished slides arguing about productivity and burnout in general, when your boss only cares about how customers get covered on Fridays. Now you're rewriting all ten.

  1. Open a new chat and send the strong version. It should stop after the plan and wait.

    Strong version: paste into any chatbot
    I need a 10 slide presentation on why our team should try a four day week. Don't write any slides yet.
    
    First give me your plan: the audience you're assuming, the whole argument in one sentence, and one line describing each slide. Then stop and wait. Only write the slides when I say "go".
  2. Correct the plan, then say go. Reply with one change, such as The audience is my boss, who cares about customer cover. Rework slides 3 and 4 around that. Then go.

    Tip
    Why it wins. A builder shows you drawings before laying bricks, because moving a wall on paper costs nothing. One line per slide takes seconds to fix. Ten finished slides take an afternoon, and the assumptions it lists are often the first thing you'd have corrected.

Your turn

  1. Take a decision you're stuck on, even a small one, and run prompt 6 on it. Write your own scoring criteria.

    Tip
    Stuck? Ask yourself what would make you regret the choice in six months. Each answer flips into a criterion. Worried about the cost? Score it for cost.
Checkpoint
You've got a scored shortlist with one clear pick, and you've corrected a plan before any of the real work was written.
If it breaksIt ignored "stop and wait" and wrote the whole thing + 1 more
It ignored "stop and wait" and wrote the whole thing
Reply: Stop. Give me only the plan, in one line per slide. Then put "Don't write it yet" at the very start of the prompt next time.
All ten options sound the same
Name the styles you want, for example one plain, one funny, one premium, one made-up word.
Still stuck? Ask the community
STEP 05 of 06~7 min

Prompts 8 and 9: Make it push back

Turn a chatbot that agrees with everything into a critic that finds the holes, and a tutor that makes you think.

Chatbots lean towards being agreeable. These two prompts give yours a different job: a critic for your plans, and a tutor that makes you think.

Prompt 8: Make it attack its own answer

Use this before you commit to an idea, plan or important draft.

  1. Send the weak version in a new chat.

    Weak version
    Is this a good business idea? A monthly subscription box of locally made dog treats.

    You'll get something like this"This has real potential! The pet market is popular and local products have a strong story." Then some encouraging tips. Lovely to read, and no use at all.

  2. Open a new chat and send the strong version. Compare how the two answers would change what you do on Monday.

    Strong version: paste into any chatbot
    Here's my business idea: a monthly subscription box of locally made dog treats.
    
    Attack it. Act like a sceptical investor who has watched fifty subscription boxes fail. Give me the 3 most likely reasons this dies, ranked, and for each one the specific evidence that would prove you right or wrong. Then give me one test I could run this week for under £50. No compliments. Under 200 words.
    An illustration of the difference. Your own reply will name its own risks.
    Tip
    Why it wins. Ask for an opinion and a chatbot leans towards being agreeable. Give it the job of finding faults and being useful now means finding faults. The evidence request keeps it from nitpicking, and the £50 test gives you something to do next.
    Tip
    Attack any answer. After any reply you're about to rely on, send: Now attack your answer. What's the weakest part, what did you assume, and what would change your mind? For an even tougher read, paste the answer into a new chat so it isn't defending its own work.

Prompt 9: Teach me, then test me

Use this when you need to understand something, not just have it explained at you.

  1. Send the weak version in a new chat.

    Weak version
    Explain compound interest.

    You'll get something like thisa clear definition, a formula and a tidy example. You nod, close the tab, and it's gone by Thursday.

  2. Open a new chat and send the strong version.

    Strong version: paste into any chatbot
    Teach me compound interest. I'm a complete beginner.
    
    Use one everyday analogy and one worked example in pounds, in under 120 words. Then ask me one question to check I've got it, and wait for my answer. If I get it wrong, don't give me the answer. Give me a hint and let me try again. When I get it right, make the next question harder.
  3. Answer its question in your own words, and don't peek.

    Tip
    Why it wins. Reading an explanation feels like learning. Having to answer shows you what you actually took in. The hint rule keeps you doing the thinking, and the harder follow-ups mean it keeps up with you.

Your turn

  1. Run prompt 8 on a plan of your own. Pick a critic who would really know, such as a tired head teacher, a landlord or the customer who never buys.

    Tip
    Stuck? Think about who could say no to this plan, or who would lose out if it went ahead. That's your critic.
Checkpoint
You've got three specific risks on a plan of yours, plus one cheap way to test them, and you've answered a chatbot's question in your own words.
If it breaksThe attack feels harsh + 1 more
The attack feels harsh
That's the point of it, and it's a chatbot being asked to play a role, so judge each risk on its evidence. Ask Which of these could I fix cheaply? to turn the list into a to-do list.
It gave me the answer instead of a hint
Reply: You gave it away. New question, and hints only this time.
Still stuck? Ask the community
STEP 06 of 06~6 min

Prompt 10: Make it write your prompts, and build your kit

Stop starting from scratch. Leave with a reusable prompt of your own and all ten in one place.

You've now used nine moves. This last one makes the chatbot do the prompt writing for you, and then you'll gather everything into a kit.

Prompt 10: Make it write your prompt

Use this for any job you do again and again, or when you've no idea how to ask.

  1. Send the weak version in a new chat.

    Weak version
    Write me a good prompt for summarising meetings.

    You'll get something like thisa one size fits all template that knows nothing about your meetings, your team or what you do with the summary.

  2. Open a new chat and send the strong version. Answer its questions, then copy the finished prompt it gives you.

    Strong version: paste into any chatbot
    Every week I want to turn messy meeting notes into a short summary and a list of actions. Write me the best possible prompt for this.
    
    First ask me up to 5 questions about anything you need to know, one at a time. Then give me the finished prompt in a code block, with [square brackets] around the parts that change each week. Then tell me what to paste into each bracket.
    Tip
    Why it wins. It already uses all the moves you've just learned, so it turns your answers into a full brief. You only do this work once, and the brackets turn the result into a form you fill in every week.

Build your kit

  1. Copy all ten templates into your notes app. Replace each [bracket] when you use one.

    Your kit: all 10 prompts
    1. JOB AND CONTEXT
    You're helping [who]. Situation: [the facts it can't know]. Task: [one thing, starting with a verb]. The goal is [what a win looks like]. Rules: [length, tone, what to avoid]. Use [square brackets] for anything you don't know.
    
    2. INTERVIEW ME
    I want [thing]. Before you start, interview me. Ask one question at a time, wait for my answer, and stop when you know enough to make this specific to me. Start with the question that would change the result the most.
    
    3. SHOW AN EXAMPLE
    Write [thing] about [topic]. Match the voice of these examples exactly: "[example 1]" "[example 2]". Output only the result.
    
    4. STICK TO THE SOURCE
    Answer using only the text below. Quote the exact line you used for each answer. If the text doesn't cover something, write "Not in the document" and don't guess. Questions: [questions]. Text: "[paste]"
    
    5. FORCE A FORMAT
    [Request]. Format: [table, list or template], with [these columns or sections], maximum [limit]. No introduction and no closing paragraph.
    
    6. OPTIONS, THEN CRITIQUE
    Context: [situation]. Give me 10 options, each noticeably different. Score each out of 5 for [criteria]. Cut to the top 3 and name the biggest weakness of each. Pick one and tell me why.
    
    7. PLAN BEFORE DOING
    I need [big piece of work]. Don't write it yet. First give me your plan: what you're assuming, the main point in one sentence, and one line per section. Then stop and wait until I say "go".
    
    8. ATTACK IT
    Here's my [idea, plan or draft]: [paste]. Attack it. Act like [the toughest critic for this]. Give me the 3 most likely ways this fails, ranked, with the evidence that would prove you right or wrong. Then one cheap test I could run this week. No compliments.
    
    9. TEACH AND TEST ME
    Teach me [topic]. I'm [your level]. Use one everyday analogy and one worked example, in under 120 words. Then ask me one question and wait. If I'm wrong, give a hint and not the answer. When I'm right, make the next one harder.
    
    10. WRITE MY PROMPT
    I regularly need to [job]. Write me the best possible prompt for this. First ask me up to 5 questions, one at a time. Then give me the finished prompt in a code block with [square brackets] around the parts that change, and tell me what goes in each.
    Tip
    Stack them. The best work usually uses a few together. Interview me, then plan before doing, then attack the draft. Each move fixes something the last one can't.
  2. Save this map with your kit.

    When an answer disappoints, find what went wrong in the grey lines and use the prompt named under it.

    The what went wrong map. Keep it near your kit.
    Tip
    Stop repeating yourself. If you paste the same background into every chat, put it in your settings. ChatGPT has custom instructions and Claude has profile preferences, and both are applied to your chats so you only write it once.

Your turn

  1. Pick the job you do most often with a chatbot and run prompt 10 on it. Test the prompt it writes on one real example, then save it to your kit.

    Tip
    Stuck? Look at your chat history and find a request you've made more than once. If you're new to this, try weekly emails, meal plans or turning notes into a to-do list.

Where to go next

Once a prompt does a job you depend on, it deserves a safety net, because a small wording or model change can quietly break it. Test your AI prompts like code shows you how to catch that before anyone else does.

Checkpoint
Your notes app has all ten templates, the map and one prompt of your own that you've used on a real example.
If it breaksThe prompt it wrote is huge + 1 more
The prompt it wrote is huge
Reply: Cut this to the shortest version that would still work, and tell me what you removed.
The saved prompt gives worse results on a new example
Paste the bad result back in and say: This prompt gave me that result. Here's what was wrong. Rewrite the prompt so it can't happen again.
Still stuck? Ask the community

Your first wins

It works. Now do something with it.

The one-line upgrade

Add this to the end of any request when you're in a hurry. It catches missing context before it can turn into a wrong answer.

Prompt
Before you answer, tell me the 3 most important things you'd need to know to do this well. If I've already told you, carry on. If not, ask me.

Check before you send

Paste any important message and find out how it lands before the other person does.

Prompt
Here's a message I'm about to send to [who]. Read it as they would. What's unclear, what could annoy them, and what would you cut? Then give me a tighter version in my voice.

Message: [paste]

Meeting notes to actions

Prompts 4 and 5 working together. Swap out any names you don't want to share first.

Prompt
Using only the notes below, give me a table with 3 columns: Action, Owner, Deadline. Write "Not stated" where the notes don't say. Under the table, list any decisions in one line each. No introduction.

Notes: [paste]

Get unstuck on a decision

Prompts 6 and 8 together, for when you've been going round in circles.

Prompt
I'm deciding between [A] and [B]. What matters most to me is [criteria]. Score each out of 5 on those criteria, recommend one, then attack your own recommendation and tell me what would change your mind.

Undo it cleanly

How to remove everything this mission installed.

There's nothing to install, so there's nothing to remove. To tidy up, delete any practice chats from your chatbot's chat history.

To stop future chats being used for training, look in your privacy settings: ChatGPT calls the setting "Improve the model for everyone", Claude calls it Model Improvement, and in Gemini it's Keep Activity. Standing instructions you added can be deleted from the same settings area you put them in.

Next mission · Use AI properlyClaude for beginners: everything it can do, in under an hour
→