Start here: a capable helper who has never met you#
Picture a note for a friend who will feed your cat while you are away. You would not write "Feed the cat." You would say which food, how much, and that the gray one bites. Your friend is smart. They just were not there when you learned all that.
An AI chat assistant, such as ChatGPT, Claude, Gemini or Copilot, is in the same spot, with one change. It has read a vast amount of general text, so it knows a lot about cats in general. It knows nothing about your cat. The message you type to it is called a prompt, and the prompt is how your situation gets in.
So an assistant has two places to draw on:
- Its training: the large body of text it learned from before it was released. This is general knowledge, such as what a picnic usually needs.
- The conversation: everything written in this chat, plus anything you paste or attach. The people who build these tools call this the context window. Anthropic's glossary describes it as a "working memory," separate from the training data.
Some assistants can also search the web, open files you connect, or remember earlier chats, if those features are on. Even then, it sees only what the feature fetched. So assume it knows nothing about your situation that you have not put in front of it.
Our running example
All through this module you are Dana. Dana coordinates the volunteers who clean up a neighborhood park, and wants to
throw them a thank-you picnic. Dana's first message to an AI is Help me plan a picnic.
Before you read on, list everything the AI would have to guess to answer Dana well.
Show the answer
How many people, and who. When and where. What the place has, such as tables or a grill. The budget. Food needs, such as vegetarians or allergies. And the big one: what Dana wants back. Ideas? A shopping list? An email? That is at least seven guesses. The AI will fill each one with something typical, and a typical picnic is not Dana's picnic.
The six parts of a strong request#
Imagine Dana asks a friend in person instead. The friend would ask a few questions before helping. Which ones? Try to sort them into groups.
Show the answer
Probably: what is it for, and who is coming? What do you already know? What can you spend? What did you like last time? Anything you do not want? Those are the six parts below.
A strong request answers, ahead of time, the questions a good helper would ask.
| Part | The question it answers | Dana's picnic |
|---|---|---|
| 1. The goal | What are you trying to get done, and why? | A thank-you picnic. I need a food plan and a shopping list. |
| 2. Context and audience | Who is it for? What is the situation? | 22 people, some kids, relaxed. Saturday 11 to 2, at a shelter with tables but no grill or power. |
| 3. The material | What real text, data or file does it need to see? | The sign-up sheet facts: 3 vegetarian, 1 nut allergy, 6 bringing a side. |
| 4. Limits | What must the answer fit inside? | $120 in total. Nothing that must stay hot. One page. |
| 5. An example | What does good look like? | Last year's list: grouped by store section, a quantity on every line. |
| 6. What you do not want | What should it avoid, and do instead? | No alcohol (park rule). No cooking on site, so make-ahead or ready-to-eat food. |
Limits are also called constraints: the edges a good answer must stay inside. You do not need all six every time. "What is the capital of Peru?" needs only a goal. The more the answer depends on your situation, the more parts you need.
Look at how part 6 is written. "No cooking" alone gives the AI nothing to aim at. Saying what to do instead gives it a target. Anthropic's guide says the same: "Tell Claude what to do instead of what not to do."
The companies agree on the basics
Here is what four AI companies' own prompting guides said when we read them in October 2026. They describe their own products, so treat this as guidance, not proof. Still, they land on the same few ingredients.
| Guide | What it says a request should carry |
|---|---|
| Google, for Gemini | A persona (a role for the AI to play), a task, context and a format. "You don't need to use all four in every prompt." |
| Microsoft, for Copilot | The goal, context, expectations and source. Only a clear goal is required. |
| OpenAI, for ChatGPT | Prompts that are "clear, specific, and provide enough context," refined over several tries. |
| Anthropic, for Claude | Treat it as "a brilliant but new employee who lacks context." Explain why, and show examples. |
Anthropic's guide offers a check that works with any assistant: "Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they'd be confused, Claude will be too." Before you send something that matters, picture a capable stranger reading it. Would they need to ask you anything? Put that answer in.
Weak or strong: predict the reply#
Look at the three pairs below. For each one, predict which reply Dana could use without another round of messages. Then say what the weak one leaves the AI to guess.
Show the answer
The strong version wins each time. Each one removes guesses and says what the reply should look like. The weak versions are not "wrong": they are fine openers if you then answer the AI's questions. They only cost you when you take the first reply as the answer.
| Weak | Strong | Left to guesswork |
|---|---|---|
Help me plan a picnic. | The full request in the box below. | Almost everything. |
Write an email to the volunteers. | Write the invitation email to our 22 volunteers. Warm and plain, under 120 words. Include the date, time and place. Give me the finished text, ready to send. | Purpose, length, tone, what must be in it, and draft or finished. |
Make it better. | The second paragraph sounds like a company memo. Rewrite only that paragraph so it sounds like a friend. Keep the rest exactly as it is. | Which part bothered you, what "better" means, what to leave alone. |
Open any free assistant. In one new chat, send Help me plan a picnic. In a second new chat, send this
word for word. Everything in it is made up.
Goal: I'm organizing a thank-you picnic for the volunteers who clean up
our neighborhood park. I need a food plan and a shopping list.
Context: 22 people, mostly adults, a few kids. A Saturday, 11 a.m. to
2 p.m., at the park's picnic shelter (reserved: 4 tables, no power, no
grill). A relaxed, friendly group, not fancy.
Material, from our sign-up sheet: 3 people are vegetarian, 1 has a nut
allergy, 6 will each bring a side dish.
Limits: $120 for everything we buy. Nothing that has to stay hot.
One page.
Example of what I like: last year's list was grouped by store section,
with a quantity on every line.
Avoid: alcohol (park rule) and anything that needs cooking at the park.
Instead, plan food we can make ahead or buy ready to eat.
Form: the finished shopping list with prices that add up, not tips
and not a template.Put the two replies side by side and look for three things:
- Guesses. In the first reply, where did it assume a headcount, a place or a budget? Did it say it was assuming?
- Fit. In the second, add up the prices yourself. They are estimates, not your store's prices. If it calls an item nut-free, that is a guess too: the package label decides. A strong request makes a good answer possible; it does not make every number right.
- Usable today. Which reply could Dana take to the store this afternoon?
Neither can know that last year the crew staged a small rebellion over potato salad. Some things you just have to say.
The strong request is longer only because it carries information. Repeating yourself or writing "this is very important" adds words, not facts. Test each line: does it tell the AI something it could not guess?
Give it the real material, and say which one#
Dana could write "some people have dietary needs", or paste the three facts from the sign-up sheet. What changes in the reply?
Show the answer
With the vague version, the AI has to invent a typical mix of needs. With the real facts, it can plan for exactly 3 vegetarians and keep nuts off the list. Real material turns a guess into a fact.
The material is the part people skip most. The rule: paste it, do not describe it. Feedback on a draft? Paste the draft. Help with numbers? Paste the rows or attach the file. A description is your memory of the thing, already shortened.
Dana's request says one guest has a nut allergy. It does not say which guest. The AI needs the fact, not the person. Leave out names, contact details and private facts about other people unless the task truly needs them. Before you paste anything from school or work, check your school's or employer's policy on AI tools.
Kestrel Exams is a small study site built by one person working with an AI assistant. Here is something that happened. Asked to "review and update the plan" for a video series, our AI assistant wrote a new plan from scratch. A detailed 23-module plan had existed for a week. It sat in the same cloud folder the assistant had opened minutes earlier, next to a note titled "read before continuing". Rule for the person: point the AI at the documents that matter. Rule for the AI: search for the plan before writing one.
Say which one
The same problem shows up with names. If you have two of something, say which one you mean.
The owner asked our AI assistant about "the progress on the photosynthesis video". It briefed him at length on a finished video. He meant a second, different video on the same topic. The clue was in the question: nobody asks about "progress" on something already finished. Rules: when two projects share a topic, name which one. When an answer does not quite fit your question, that mismatch is a clue the AI is thinking of something else.
For Dana: "the November thank-you picnic", once the spring picnic has come up in the same chat.
Ask for the form you want, or for questions first#
Dana types: "Help me with the invitation email." Name three different things an AI could reasonably hand back.
Show the answer
Tips on writing invitations. A template with blanks such as [DATE]. A finished email, ready to send. A list of subject lines. All of them count as "help". Only one is what Dana wanted, and the request did not say which.
"Help with" is vague. Say what you want to be holding when the reply ends.
The owner asked our AI assistant for some prompts he needed. It handed over a reusable prompt template plus a list to paste into it, which left the assembly to him. He wanted the finished prompts. Rule: say what form you want the result in.
| If you want | Say something like |
|---|---|
| The finished thing | Give me the finished email, ready to send. |
| Something reusable | A template with blanks I can fill in for future events. |
| A shape and a size | A table: item, quantity, price. One page. |
| To learn, not an answer | Don't solve it. Ask me questions that lead me to the answer. |
An example sets the form well. Anthropic's guide calls examples "one of the most reliable ways" to steer format, tone and structure. Dana's line about last year's list is an example doing that job.
Not sure what to include? Ask it to ask you
If you have never done the task before, you may not know what a good request needs. Turn it around. Give the AI your goal and ask it to question you first. A clarifying question fills a gap before the work starts, the way a good helper would.
In a new chat, send this:
I'm planning a thank-you picnic for about 20 volunteers. Before you
suggest anything, ask me up to five questions whose answers would most
change your plan. Number them, ask them all at once, and wait.Answer with made-up details. Which of the six parts did its questions cover? Did it ask something you would not have thought to mention? Add that one to your own requests next time.
Steer it: follow-ups, steps and roles#
The first invitation draft is fine, except the second paragraph sounds like a company memo. Dana types "Make it better." What might change that Dana already liked?
Show the answer
Anything at all. The AI does not know which part bothered Dana, so it may rewrite the whole email, including the parts that worked. A good follow-up points at the part, says what is wrong, and says what to keep.
The first reply is rarely the last. OpenAI's guide says "Prompt engineering often requires an iterative approach." Prompt engineering just means writing and refining prompts with care. A follow-up, your next message in the same chat, does most of the steering:
Shorter: under 80 words. Keep the date, time and place.Check the total. Show each price and the sum, and bring it to $120 or less.You assumed there is a grill. There isn't. Redo the food plan.
When a reply rests on a wrong assumption, name the assumption. Do not just ask for "another version".
Break big jobs into steps
The whole picnic is too much for one request. Dana sends it in steps: the food plan, then the shopping list, then the invitation, then a checklist for the day. Each step can be checked before the next builds on it. A plan that quietly assumes a grill gets caught in step one, not baked into the list and the email.
Roles: a shortcut, not a spell
A prompt starts "You are a world-class event planner." What does that tell the AI that Dana's six-part request does not?
Show the answer
Almost nothing about the picnic. It hints at tone and the kind of knowledge to use, but adds no facts.
Some prompts start "Act as an experienced event planner." A role like that is a shortcut for context: it tells the AI which kind of knowledge and tone to lean on. Anthropic's guide says a role "focuses Claude's behavior and tone." Useful. But a role adds no facts about your situation, and it does not make the answer more correct.
No phrase works like a password. "You are a world-class expert" does not give the AI facts it lacks or make it check its work. When a role helps, it is because it carried context. "You are helping a volunteer group with no paid staff and a small budget" says far more than "you are an expert".
When you need to choose: ask for a recommendation#
Dana is torn between the park shelter and a volunteer's big backyard. Dana could ask "Which is better?" What will a reply to that probably lack?
Show the answer
Reasons tied to Dana's situation, and any sign of what the AI assumed. "Which is better?" invites a one-word verdict or a vague list of pros and cons. Neither shows the assumptions underneath.
When you have to decide, ask for a comparison on the points that matter, a recommendation, and the reasons. Then add one question that does a lot of work: what would change your mind?
I need to choose a place for the volunteer picnic: the park shelter
(reserved, roofed, 4 tables, public restrooms, gets hot by midday) or a
volunteer's backyard (shady trees, no roof, one bathroom, street
parking only). Compare them on weather, bathrooms and parking.
Recommend one, give your reasons in three sentences, and tell me which
fact would change your answer.The last line drags hidden assumptions into the open. "If rain is likely, keep the roofed shelter. If it will be very hot, choose the shady backyard." That tells you to check the forecast.
The choice stays with you. The AI does not know that the backyard owner's dog dislikes crowds. Treat the recommendation as one well-argued opinion, then decide.
Standing instructions, and when to start fresh#
After a few weeks, Dana is typing the same lines in every chat: "give me the finished thing", "plain words", "ask me if something is missing". What could Dana do?
Show the answer
Write those lines once, as standing instructions, and reuse them. It saves typing, and it sets the AI's defaults to match how Dana works.
Standing instructions are a short note about you and how you like answers, given at the start of every conversation. Here is Dana's:
About me: I coordinate volunteers for a small park cleanup group.
No paid staff, small budgets, an informal crowd.
- Give me the finished thing, not tips or a template, unless I ask.
- Plain words, short sentences.
- If my request is missing something important, ask me up to three
questions first.
- When I'm choosing, recommend one option and give your reasons.
- If you are unsure of a fact, say so instead of guessing.If your assistant has a settings area for standing instructions (each product names it differently), put it there. If not, keep it in a note and paste it in. Instructions shape replies; they do not guarantee them, so you still check what comes back.
Our free starter prompt turns standing instructions into a full memory system. Upload it to Claude or ChatGPT and it walks you through setup. See how it works.
When to start a fresh conversation
Everything in a chat, wrong turns included, is part of what the AI reads before each reply. More is not always better. Anthropic's guide to context windows says more context "isn't automatically better": as the amount of text grows, accuracy and recall drop. Three signs say it is time for a fresh chat:
- You are switching to an unrelated task.
- It keeps bringing back something you already corrected.
- The chat is so long you could not easily sum up what was decided.
Starting fresh does not mean starting from nothing. Bring a short summary with you.
At the end of any planning chat (made-up details only), send:
Summarize what we have decided as a short note I can paste into a new
chat: decisions, open questions, and anything I asked you to avoid.
No more than 10 lines.Read the summary before you use it. Is every decision one you made? Is anything you corrected written the old way? Fix it by hand, then paste it into a new chat. You have just written a strong request, mostly for free.
What was new in this module#
- An AI assistant draws on its training (general knowledge) and the conversation (your situation). Your situation gets in only through what you write, paste or attach.
- A strong request has six parts: the goal, context and audience, the real material, limits, an example, and what you do not want. Pair each "don't" with what to do instead.
- The colleague test: if a capable stranger would need to ask you something, put the answer in.
- Paste the material, do not describe it. Give the facts the task needs, not other people's private details.
- When two things share a name or topic, say which one. A reply that does not fit your question is a clue.
- Name the form you want back, or ask the AI to ask you questions first.
- Steer with specific follow-ups: point at the part, say what is wrong, say what to keep.
- Break big jobs into steps you can check one at a time.
- A role is a shortcut for context, not a spell.
- To decide, ask for a recommendation, its reasons and what would change it. Then decide yourself.
- Write standing instructions once. When a chat drifts, start fresh with a checked summary.
It knows the world, not your situation. Put the goal, the context, the real material, the limits and the form you want into the request, then steer with specific follow-ups.
Key terms#
- Prompt
- The message you send to an AI assistant: your request, plus anything you paste in.
- Training
- The large body of text an AI learned from before release. It gives general knowledge, not knowledge of your situation.
- Context window
- The text a model can look back on while it writes: this conversation plus anything pasted or attached. A working memory, separate from training.
- Constraint
- A limit the answer must fit inside, such as a word count, a budget, a tone or a format.
- Clarifying question
- A question that fills a gap before the work starts. You can ask the AI to ask you these first.
- Follow-up
- Your next message in the same chat, used to steer: shorter, a different tone, a corrected assumption.
- Role
- An instruction such as "act as an event planner". A shortcut for context and tone; it adds no facts and no accuracy.
- Standing instructions
- A short note about you and how you like answers, reused at the start of every conversation.
Check yourself#
Answer each one out loud before you open it. Reading the answer without trying first feels like learning and is not.
1. A classmate types "Summarize the article I read yesterday" and pastes nothing. What will go wrong, and what is the fix?
The AI has never seen the article. At best it says so. At worst it guesses which article was meant. The fix: paste the text or attach the file, and say what the summary is for and how long it should be.
2. Your prompt says "Don't make it too formal", and the reply is still stiff. What would you change?
Say what you want instead: "write it like a text to a friend, with short sentences". Better still, paste a short example of the tone you like.
3. You asked for help with a study plan and got five tips about planning. Do you start over? What do you send?
No need to start over: the request never named the form. Send a follow-up such as "Give me the finished plan as a table: day, topic, minutes, for two weeks." Then check it against your real schedule.
4. A coworker says the secret to good answers is starting every prompt with "You are a world-class expert." How do you judge that?
A role can focus the tone and the kind of knowledge, so it is not useless. But it adds no facts and no accuracy. Describing the actual situation does more work than any title.
5. After a long chat about a group project plan, the AI keeps bringing back a deadline you corrected twice. What do you do?
Ask for a short summary of what was decided. Check it and fix the deadline by hand. Then start a fresh chat with the corrected summary as the first message.
I can#
Tick each one only when you could do it tomorrow without this page open.
- I can explain where an AI assistant's knowledge comes from, and why it needs the details of my situation.
- I can turn a one-line request into a strong one using the goal, context, material, limits, an example and what to avoid.
- I can name the form I want back, or ask the AI to ask me questions first.
- I can steer a reply with a specific follow-up instead of "make it better".
- I can ask for a recommendation with reasons and still make the decision myself.
- I can write my own standing instructions and tell when to start a fresh conversation.
Frequently asked questions#
How do I write a good prompt for an AI chatbot?
Say what you want done and why, who it is for, and the situation. Paste the real material. Add limits, an example of good output, and what to avoid. Say what form you want back. A simple question needs only the goal.
Are there magic words that make AI answers better?
No phrase works like a password. Politeness is fine, and a role can set the tone. What changes the answer is information: the facts of your situation, the material, and a clear description of the result you want.
Why does the AI give me generic answers?
Usually because the request left it to guess. It fills each gap with something typical, and typical reads as generic. Add the missing details, or ask it to ask you questions first.
Should I ask the AI for a template or for the finished thing?
Ask for whichever you will actually use. Sending an email today? Ask for the finished email. Reusing the same shape often? Ask for a template. If you do not say, you may get either one, or a list of tips.
When should I start a new chat instead of continuing?
When you switch to an unrelated task, when the AI keeps bringing back something you corrected, or when the chat is too long to sum up. Ask for a short summary first, check it, and paste it into the new chat.
This course is free, with no ads and no account. If it helped you, you can help keep it that way.
Support Kestrel Exams →Sources#
Anthropic. Prompting best practices. Claude Platform Docs, undated, read October 21, 2026. platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices
Anthropic. Glossary ("Context window"). Claude Platform Docs, undated, read October 21, 2026. platform.claude.com/docs/en/about-claude/glossary
Anthropic. Context windows. Claude Platform Docs, undated, read October 21, 2026. platform.claude.com/docs/en/build-with-claude/context-windows
OpenAI. Prompt engineering best practices for ChatGPT. OpenAI Help Center, updated October 2026, read October 21, 2026. help.openai.com/en/articles/10032626
Google. Writing effective AI prompts for business. Gemini for Workspace, undated, read October 21, 2026. workspace.google.com/resources/ai/writing-effective-prompts
Microsoft. Learn about Copilot prompts. Microsoft Support, undated, read October 21, 2026. support.microsoft.com/en-us/topic/learn-about-copilot-prompts