The simplest definition comes first, and then we make it useful.

A prompt is the input you give to an AI model to tell it what you want. It can be a question, an instruction, a piece of text to work on, an example to copy, or any mixture of these. When you type a message to Claude and press send, that message is your prompt.

The word “prompt” comes from the idea of prompting, as when a teacher gives a student a starting line, or a stage assistant whispers the first words to an actor. The prompt gets the performance started. The AI’s reply is called the response, or sometimes the output. So every exchange has two halves: you write the prompt, and the model writes the response.

A Prompt Is More Than a Question

Many beginners think a prompt must be a question. It does not. All of the following are prompts: “Explain photosynthesis to a ten-year-old.” “Rewrite this email so it sounds friendlier.” “Here is my essay. Tell me the weakest paragraph.” “Give me five names for a bakery.” Each one tells Claude what to do, and each one gets a response.

People have always had to tell computers what to do. What changed is how we say it.

In the early days of computing, you gave instructions in strict, special languages. A single misplaced symbol could break everything, because the computer understood only the exact commands it had been built to expect. Talking to a computer meant learning its language.

Later, search engines let ordinary people type short keywords and get results. That was friendlier, but it still taught people to shrink their thoughts into a few words, because that was what the machine could match. Voice assistants then allowed spoken commands, though usually for a short list of simple tasks.

Large language models changed the situation. A large language model, often called an LLM, is a program trained on a huge amount of text so that it can read and write human language. Claude, made by the company Anthropic, is one such model family. Because these models were trained on language itself, you can talk to them in ordinary sentences, describe a messy situation, and receive a thoughtful reply. The computer finally meets people halfway.

Twelve short definitions that make the rest of this guide easy to follow.

A strong prompt is usually built from a handful of simple parts, and you can pick the ones you need.

Not every prompt needs every part. A quick question may need only one. But when a task matters, thinking through these six parts will almost always improve the result. The diagram shows how they fit together, and the sections below explain each one.

flowchart TD
    A["A strong prompt"] --> B["1. Goal
What do you want done?"]
    A --> C["2. Context
What is the background?"]
    A --> D["3. Audience
Who is it for?"]
    A --> E["4. Examples
What does good look like?"]
    A --> F["5. Constraints
What limits apply?"]
    A --> G["6. Format
How should it be shaped?"]
    

Understanding what happens inside makes the advice in this guide feel logical instead of arbitrary.

Claude does not look up an answer from a stored list. It writes the response from scratch, one token at a time. First it reads your prompt, which is split into tokens and turned into numbers. These numbers pass through many layers of the model. A mechanism called attention lets the model work out which parts of the prompt matter most for the next piece of the reply. Then the model scores every possible next token, picks one, adds it to the reply, and repeats until the answer is complete.

flowchart LR
    P["Your prompt"] --> T["Split into tokens"]
    T --> N["Turn tokens into numbers"]
    N --> L["Layers of attention weigh what matters"]
    L --> X["Scores for the next token"]
    X --> S["Choose one token"]
    S --> O["Add it to the response"]
    O --> L
    

Why the Prompt Matters So Much

Everything Claude writes is shaped by the prompt in front of it. The prompt sets the direction for the very first token, and each later token builds on the ones before it. If the prompt is vague, many different directions look equally likely, and the reply tends toward something generic. If the prompt is specific, fewer directions fit, and the reply lands closer to what you meant.

Claude Does Not Read Your Mind

Claude only has the words you give it, plus its training and any earlier messages in the chat. It cannot see your screen, your project, or your intentions unless you describe them or share them. This sounds obvious, but it explains most disappointing answers. The model did not fail to understand. It was simply never told.

Your prompt is not the only thing Claude sees, and knowing the rest explains many surprises.

When you press send in a Claude product, several pieces of text are put together before they reach the model. Your latest message is only one of them. The picture below shows the order of events, and the steps after it describe each part.

sequenceDiagram
    participant U as You
    participant A as The app
    participant C as Claude model
    U->>A: Type your prompt
    A->>A: Add the system prompt and earlier messages
    A->>C: Send everything together
    C->>C: Read all the tokens and plan the reply
    C-->>A: Stream the response in small pieces
    A-->>U: Show the response as it appears
    

What This Means for You

Because the whole conversation is sent each time, a very long chat means a very long prompt for the model to read. Old, unrelated material may distract it, and eventually the history may approach the limit of the context window. This is why starting a fresh chat for a new topic is a simple habit that often improves answers.

Prompts come in many shapes. Recognizing them helps you pick the right one.

You do not need to memorize categories, but seeing the common kinds makes it easier to choose a shape for your own request. Most real prompts are a mixture of two or three of these.

If you remember only one rule about prompts, make it this one.

The single most effective habit is being specific. A vague prompt invites the model to guess, and guesses produce average results. A specific prompt narrows the target. Being specific does not mean being long. It means saying the important things clearly.

Replace Vague Words With Concrete Ones

Say What to Do, Not Only What to Avoid

It is natural to write “do not be boring” or “do not use big words”. Such instructions are helpful, but they are even better when paired with a positive description: “use short sentences and everyday words”. Telling Claude what you want gives it something to aim for, instead of only something to avoid.

Give the Reason When It Matters

Adding a short reason helps Claude understand the spirit of your instruction. “Keep it brief because it will be read on a phone” is more useful than “Keep it brief”, because Claude can then make sensible choices you did not spell out, such as using short paragraphs.

The same words can deserve very different answers depending on your situation.

Consider the prompt “How should I save money?” A student, a retiree, and a small business owner would each need a different answer. Without context, Claude can only give general advice that fits nobody especially well. With a few lines of background, the same question becomes a request for exactly the advice you need.

What Counts as Useful Context

Without Context and With Context

Models are excellent at spotting patterns, so a sample is often worth a paragraph of explanation.

Describing a style in words is surprisingly hard. Words like “friendly but professional” mean different things to different people. An example removes the guesswork. When you paste a sample of what you like, Claude can see the length, the tone, the structure, and the little habits of the writing, and then follow them.

Giving one or more examples inside a prompt is often called few-shot prompting. A “shot” is just an example. Zero examples is a “zero-shot” prompt, one example is “one-shot”, and a few examples is “few-shot”. You do not need to remember the names. You only need to remember the idea: show what good looks like.

How to Use Examples Well

A great answer in the wrong shape is still hard to use.

Claude chooses a format on its own when you do not say, and it usually makes a sensible guess. But if you need something specific, such as a table to paste into a document or a short reply for a text message, say so. Format instructions are among the easiest to give and the most dependable to follow.

Common Formats You Can Ask For

Length

Be direct about length. “In two sentences”, “about 300 words”, and “a one-page summary” all help. If the first answer is too long, simply say “shorter, please” or “cut it in half”. If it is too brief, say “expand the second point with an example”. Length is cheap to adjust.

Tone

Tone words work well too: formal, warm, playful, calm, confident, plain. You can also describe the situation: “This is for a customer who is annoyed about a late delivery”. Claude will then choose a suitable tone without a long list of rules.

For problems that need careful thought, a small change in the prompt can make a big difference.

Some tasks are not simple lookups. They involve several stages: working out a budget, comparing options, solving a puzzle, or planning a project. For such tasks, asking Claude to reason in steps tends to give better results than asking only for the final answer. The model is built to produce text one piece at a time, so letting it write out intermediate steps gives it room to work things through.

Ways to Ask

Some versions of Claude can also spend extra time reasoning before they reply, which helps with math, logic, and complex planning. This takes a little longer, which is the trade-off. For a simple request, you do not need it.

Sometimes the best prompt starts with who should be speaking.

A role prompt asks Claude to take a particular viewpoint: “Act as a patient math tutor”, “Respond like an experienced editor”, or “Pretend you are a hiring manager reviewing my resume”. This can help Claude choose the right vocabulary, the right level of detail, and the right kind of feedback.

When Roles Help

Roles are useful for practice and perspective. You can rehearse a job interview with an interviewer role, test a presentation with a skeptical audience, or ask for feedback from the viewpoint of a customer. They can also set a helpful tone, such as a gentle tutor for a nervous learner.

Roles Are a Lens, Not Magic

A role does not give Claude new knowledge or new abilities. It focuses how the existing knowledge is presented. Saying “act as an expert lawyer” does not turn Claude into a licensed lawyer, and the answer should not be treated as legal advice. Pairing a role with clear details about your situation works better than a role alone.

Your first prompt does not have to be perfect. Most good results come from a few rounds.

Many beginners treat a prompt like a single throw: if it misses, they give up. Experienced users treat it like a conversation. They send a first prompt, look at the response, and then say what to change. Each follow-up becomes part of the context, so the conversation steadily homes in on what you want.

flowchart LR
    A["Write a clear first prompt"] --> B["Claude responds"]
    B --> C["You review the response"]
    C --> D{"Is it what you need?"}
    D -->|"Yes"| E["Use it and check key facts"]
    D -->|"No"| F["Say exactly what to change"]
    F --> B
    

How to Give Good Feedback

When your prompt includes a lot of text, a little organization goes a long way.

One of Claude’s strengths is reading long material. You can attach or paste an article, a report, a contract draft, meeting notes, or a pile of customer feedback and ask questions about it. A larger context window means more fits in a single prompt. But long prompts need good structure, so the model knows which part is which.

Keep Instructions and Material Separate

A common mistake is to paste a long document and then add a small instruction in the middle of it. Claude may not be sure where the material ends and the task begins. A better approach is to introduce the material clearly, for example by saying “Here is the report”, then paste it, and then place your question in a separate, clearly marked part. Anthropic’s prompting guidance suggests using clear labels or markers to separate parts of a long prompt, and many people find it improves results.

Tell Claude What to Focus On

If only part of a document matters, say which part. “Focus on the pricing section” saves effort and reduces distraction. For many documents, you can ask for a particular type of output, such as a list of action items with owners, or the five riskiest statements.

Ask for Evidence

When working with a document, ask Claude to point to where in the text each claim comes from. “Quote the sentence that supports each point” helps you verify answers quickly and makes it harder for unsupported details to slip in.

Behind many assistants there are two kinds of instruction, written by two different people.

The user prompt is what you type. The system prompt is a hidden set of instructions written by whoever built the product you are using. It tells the assistant who it is, how to behave, what tone to use, and what it should or should not do in that setting. A company’s support helper might have a system prompt telling it to be polite, use only the company’s help articles, and pass difficult cases to a human.

flowchart TD
    A["System prompt
Written by the product builder"] --> C["What Claude reads"]
    B["User prompt
Written by you"] --> C
    D["Conversation history"] --> C
    C --> E["Claude writes a response that fits all of them"]
    

Why This Matters to You

If Claude behaves differently in two different products, the difference is often in the system prompt, not the model. A product built for children may keep answers simple and gentle. A tool for programmers may keep answers short and technical. When you write your own prompt, you are working within whatever frame the product has set.

If You Build With Claude

People who build their own products with Claude write the system prompt themselves. Good system prompts are clear, describe the role and the audience, explain the reasons behind rules, and describe how to handle unusual cases. They benefit from the same habits as ordinary prompts: be specific, give context, and test the result.

Some prompts do not only ask for words. They ask for actions.

In many Claude products, the assistant can use tools. It might search the web, read a file you attach, run a calculation, or look into an app you have connected. When tools are available, your prompt can ask for outcomes, not just text: “Find the latest opening hours for these three museums and put them in a table.” Claude decides which tools to use, uses them, and writes the answer from what it finds.

How Prompting Changes With Tools

Prompting is powerful and easy to start, but it has limits worth understanding.

The Trade-offs You Will Feel

Whatever you put in a prompt, and whatever the model reads, deserves a moment of care.

Privacy: What Goes Into a Prompt

A prompt is information you share with a service. A good rule is to leave out anything you would not want forwarded: passwords, one-time codes, card numbers, government ID numbers, and sensitive details about other people. If you need help with a document that holds private data, replace real names and numbers with placeholders before you paste it. Also review your product’s privacy settings and optional features, such as saved memory, and switch them on or off as you prefer.

Prompt Injection: Instructions Hidden in Content

When Claude reads a web page, an email, or a document, that content can contain text that looks like instructions aimed at the assistant. Someone with bad intentions hopes Claude will treat those words as orders from you. This trick is called prompt injection. Claude is trained to treat such text as material to read, not commands to obey, and products add further protections. Still, the safest habit is to approve important actions yourself, such as sending messages or buying things.

A good prompt is not only written. It is tested, checked, and kept.

When engineers run a system, they watch it to notice problems early. You can do a light version of this with your own prompts. The idea is simple: look at what came back, decide whether it is good, and learn from the gap. Over time, this habit turns scattered tries into reliable skills.

Check the Response

Keep a Small Prompt Library

When a prompt works well, save it. A simple notes file with your best prompts, each with a line about what it is for, saves hours. For prompts you reuse, mark the parts you change each time, such as the topic or the audience, so that you can swap them quickly.

Test With a Few Variations

If a prompt matters, try it with two or three different inputs before you trust it. A prompt that works for one sample may fail on another. For people building products, teams do this on a larger scale with a fixed set of test questions that they run again after every change, to see whether quality rose or fell.

A pattern is a habit that works again and again. An anti-pattern looks tempting and keeps causing trouble.

Patterns That Work

Anti-patterns to Avoid

Watching a prompt improve, step by step, is the best way to see why each part matters.

Let us follow one request through five versions. Imagine you want help with a message to a neighbor about noisy late-night music. Each round adds one idea from this guide.

Fill-in-the-blank patterns save time. Replace the parts in brackets with your own details.

Once you know the parts of a prompt, you can make simple templates for tasks you do often. Below are recipes in plain sentences, with the changeable bits described in brackets. Copy the idea, adapt the wording, and keep the ones that work in your prompt library.

Everyday scenes show how the ideas combine in practice.

Several popular ideas about prompting are only half true.

Why These Myths Stick

When a tool seems to respond to words, people look for the perfect words. Online posts sometimes share a phrase that worked once and present it as a trick. But models respond to meaning, not to magic spells. What reliably helps is giving the model the information and direction it needs, which is exactly what the parts of a prompt are for.

A short checklist for today, and the slips to watch for.

Common Mistakes

Quick answers to what beginners ask most.

If you remember only a few ideas from this guide, make it these.