A prompt is the instruction, question, or text you give an AI tool, such as ChatGPT, Claude, or Gemini, so it can generate a response. It can be a single line ("summarize this report") or a full brief with a role, context, a task, a format, examples, and constraints. The clearer the prompt, the more useful the answer you get back.
Most people write prompts that are far too short. Google's Prompting guide 101 reports that the most effective prompts average around 21 words with relevant context, while the prompts people actually try are usually fewer than nine words. This guide shows you how to close that gap.
What is a prompt in AI? A plain-English definition
A language model doesn't know your company, your customer, or what happened in yesterday's meeting. It only works with the text it receives. That text is the prompt, and anything that isn't in it, the model has to guess.
Anthropic's prompting best practices suggest thinking of the AI as a brilliant but new employee who doesn't know your norms or workflows. Ask that person to "put together a report" and you'll get something generic. Tell them who it's for and what decision it needs to support, and the result gets much better.
One nuance: what you type isn't the only prompt. AI apps add their own instructions behind the scenes, known as a system prompt. OpenAI's prompt engineering guide explains that messages carry roles with different levels of authority (developer, user, and assistant) and that developer instructions take priority over user messages. That's why the same prompt can produce different answers in two different apps. And when you attach a document or an image, that file becomes part of the prompt too.
What a prompt is for
The prompt is your only channel for telling the AI what you need. At work, you'll use prompts to:
- Generate text: emails, posts, proposals, or scripts.
- Transform text: summarize, translate, change the tone, or shorten.
- Analyze information: pull data from a document, compare options, or sort customer feedback.
- Think more clearly: brainstorm, prep for a meeting, or stress-test a plan.
- Automate small tasks: write spreadsheet formulas or understand a code error.
- Create images: describe the scene, style, and composition you want.
In every case, the quality of the output depends on the quality of the request.
The parts of a prompt
Not every prompt needs every part, but knowing them helps you spot what's missing when an answer falls flat. Google's guide to writing effective prompts names four areas: persona, task, context, and format. OpenAI's and Anthropic's guides add two more: examples and rules or constraints.
| Part | What it adds | Example |
|---|---|---|
| Role (persona) | Point of view, vocabulary, and depth | "Act as a tax advisor for freelancers" |
| Context | The situation, audience, and facts the AI doesn't have | "My client is a dental practice that wants more online bookings" |
| Task | What it needs to do, starting with a verb | "Write three subject lines for our fall campaign" |
| Format | How you want the answer delivered | "A two-column table, no intro" |
| Examples | A model of the output you expect | "Here's how we write our subject lines: …" |
| Constraints | Limits and rules to follow | "No jargon; if you're missing information, ask me" |
The task is the one part you can't skip. Google's guide says to always include a verb or command and calls it the most important component of a prompt. For ready-made templates built on this structure, see our collection of ChatGPT prompts for work.
Types of prompts
Day to day, you'll run into four basic types, and you can combine them.
Zero-shot: a direct instruction with no examples
You ask for something without showing an example: "Translate this paragraph into Spanish with a formal tone." It works well for common tasks the model already handles, like translating, summarizing, or writing a first draft.
Few-shot: with examples
You include sample inputs and outputs so the AI can copy the pattern. The term took off with the paper "Language Models are Few-Shot Learners", published in 2020 by OpenAI researchers, which showed that a large model could handle new tasks from just a few examples inside the prompt. The Gemini API prompt design guide recommends always including examples, and Anthropic suggests 3 to 5 for best results in its best practices guide. Use it when the format or tone has to be exact.
Role prompting
You give the AI a specific role: "You're a high school math teacher" or "Act as a recruiter." The role shapes vocabulary, detail, and focus, and Anthropic notes in its guide that even a single sentence of role-setting changes the model's behavior and tone. Pair it with context: a role without details about your situation still produces generic answers.
Chain of thought: step-by-step reasoning
You ask the AI to reason step by step before giving its final answer. The technique was described in the 2022 paper "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models", and another study that same year showed that simply adding "Let's think step by step" improved results on arithmetic and logic problems. It helps with calculations and decisions that weigh several factors.
Newer reasoning models change the approach. OpenAI's guide explains that reasoning models perform better with high-level guidance, while GPT models benefit from very precise instructions. With a reasoning model, describe the goal clearly and let it work out the steps. For long tasks, split the work into a chain of prompts, as the Gemini guide recommends: first extract the data, then summarize it, then draft the email.
Prompt examples: weak vs. strong
Here are eight everyday tasks, with the version most people type and a better one:
| Task | Weak prompt | Strong prompt | What changed |
|---|---|---|---|
| Customer email | "Write an email to a customer" | "You're a customer support lead at an online store. Write a short email to a customer whose order is delayed: apologize, give the new delivery date, and end with a question. Warm tone." | Role, context, and tone |
| Meeting recap | "Summarize this" | "Summarize the notes below in five bullets: decisions, owners, and deadlines. If something is missing, write 'TBD.'" | Format and a rule for gaps |
| Social ideas | "Give me Instagram ideas" | "Suggest ten Instagram post ideas for a neighborhood bakery that wants more morning customers. Table with idea, format, and a one-line caption." | Audience, goal, and format |
| Spreadsheet | "Help me with Excel" | "Column A has dates and column B has amounts. Give me a formula that sums only March amounts and explain it in two sentences." | Specific data |
| Learning a concept | "What is inflation?" | "Explain inflation like I'm 15, using a weekly grocery run as the example, in under 150 words." | Level, example, and length |
| Sorting reviews | "Are these reviews good?" | "Label each review as positive, negative, or neutral. Example: 'Arrived late but works great' = neutral. Return a table." | Clear criteria and one example |
| Resume feedback | "Fix my resume" | "Act as a marketing recruiter. Review this resume for a junior analyst role and give me the three changes with the biggest impact. Don't invent experience." | Role, goal, and a constraint |
| Code error | "Fix my code" | "This Python script throws KeyError: 'email' (code and error pasted below). Explain the cause in one sentence and give me the smallest fix." | Technical context |
Strong prompts aren't longer for the sake of it: every added sentence answers a question the AI would otherwise have to guess. To see these ideas applied inside the chat interface, read our guide on how to use ChatGPT.
Common prompt mistakes
- Asking without context. "Write a post about sales" doesn't say who it's for or what it should achieve, so the AI fills the gaps with the most generic option.
- Stacking several tasks in one message. Ask it to analyze, summarize, and draft at once, and some part will come out weaker.
- Only saying what you don't want. Anthropic's guide recommends telling the model what to do instead of what not to do: rather than "no jargon," write "explain it for someone who has never used a spreadsheet."
- Skipping the format. You get long paragraphs when you wanted a table you could paste straight into a doc.
- Accepting the first answer. Google's guide advises reviewing AI output for clarity, relevance, and accuracy before you use it.
- Asking for facts it doesn't have. Without sources, it can make up numbers. Provide the data yourself or ask it to flag missing information.
- Pasting confidential information. Before you share customer data, check what your company's AI policy allows.
How to write a prompt, step by step
- Decide what you'll do with the answer. A draft for yourself needs something different than a message for a client.
- Write the task with a verb. Draft, summarize, compare, classify, explain.
- Add the context only you have. Audience, goal, and data. If you paste a long document, put it before your instructions.
- Assign a role if it changes the angle. Recruiter, teacher, editor.
- Ask for a format and a length. Table, list, email, number of bullets.
- Add an example if style matters. One good example saves several rounds of edits.
- Set constraints. What it shouldn't invent and what to do if information is missing.
- Test and refine. The Gemini guide points out that prompt design is iterative. Use Anthropic's golden rule: if a colleague with no context would be confused by your prompt, the AI will be too.
Here's a prompt that follows those steps:
You're the training lead at a logistics company. Create an outline for a 30-minute session that explains to warehouse supervisors what a prompt is and how to use one to write incident reports. The group has no technical background. Organize it into three timed blocks, end with a hands-on exercise, and keep the language jargon-free.
If you'd rather practice with guided exercises, our free prompt engineering course is interactive and takes about 5 hours.
Prompt vs. prompt engineering
A prompt is a single instruction. Prompt engineering is the process of designing, testing, and improving those instructions so they deliver good results consistently. OpenAI's guide defines it as the process of writing effective instructions for a model so that it consistently generates content that meets your requirements.
| Prompt | Prompt engineering | |
|---|---|---|
| What it is | One instruction | A method for designing and improving instructions |
| Scope | One answer | Repeatable results across many cases |
| Who does it | Anyone who uses AI | Teams building AI into workflows or products |
| How it's measured | Does this answer work for me? | Success criteria and tests across many cases |
The key difference is measurement. Anthropic's prompt engineering overview says that before optimizing a prompt, you should have clear success criteria and a way to test against them. You don't need to become a prompt engineer to use AI at work, but you do need the basics: what to ask for, how to ask, and how to check what you get back.
To see which other tools fit your work, browse our map of AI tools.
Your next step: turn repetitive tasks into prompts
This week, pick three tasks you repeat often, like recapping meetings or answering customer questions. Write a prompt for each one using the parts above, refine it, and save the version that works. That's the start of your own prompt library.
If you want structure and support, 4Geeks offers AI Fluency, a 4-week program for using AI at work without coding, with 1:1 mentorship, a 24/7 AI tutor (Rigobot), and a certificate. If you want to go further, our Applied AI Course runs 6 weeks for working professionals and includes a certificate recognized by the Florida Department of Education. Still weighing options? Our roundup of the best AI courses for beginners can help you narrow them down, and when you're ready to decide, you can compare our programs side by side.