Imagine asking a neighbor, “Could you pick up some things for dinner?”
They return with bread, chicken, and a chocolate cake.
You were planning vegetable soup.
Your neighbor was willing to help. You left out the part that mattered.
Something similar happens when we use AI. We type “Write an email” or “Help me improve my business,” then wonder why the answer sounds generic.
The system has been given a large job and very little direction.
Prompt engineering is the practice of making that direction clearer. A prompt is the request and information you give an AI tool. Improving it means explaining the job, supplying the right material, and describing what a useful answer looks like.
You can do that in ordinary language.
Begin with a real task
Suppose a community garden needs volunteers for a Saturday cleanup. You want a short email that makes helping sound welcoming and manageable.
You could type:
Write an email asking for volunteers.
But the tool does not know who will read it, what the work involves, or how someone should respond. It may fill the gaps with cheerful language that tells readers very little.
For this fictional event, a more useful request would be:
Write a friendly email inviting neighborhood residents to help clean up our community garden. The event is Saturday, October 10, from 9 to 11 a.m. Volunteers will pull weeds and spread mulch. No gardening experience is needed. Ask readers to reply to the email if they want to help. Keep the body under 150 words and include a subject line. Use everyday language. Do not add details beyond the ones provided.
Now the task has shape.
The email has an audience, a purpose, confirmed details, and a clear next step. You can also spot what is missing: the garden’s location should be supplied before the invitation goes out.
The following six techniques help you give AI that kind of direction.
1. Explain the task and the context
Start with what you want done. Then add the background needed to do it well.
For the garden email, “The readers are beginners who may worry they cannot help” is useful context. It suggests that the invitation should explain how simple the work is.
You can also give the AI a role:
Act as an editor helping a small community organization write clearly.
A role helps set the perspective. It does not give the system new facts or guarantee expertise. “You are the world’s greatest editor” matters less than explaining who the message is for and what it needs to accomplish.
Try this order: task, audience, relevant background, desired result.
If you are unsure what information matters, ask:
Before drafting, identify any missing details that would prevent a reader from acting on this invitation.
That gives the conversation a useful first step.
2. Ask for evidence you can check
For a simple writing task, a clear brief may be enough. For a decision or analysis, ask the tool to connect its answer to the material you provided.
Suppose you want help assessing a return request. Supply the actual policy and the customer’s message, then ask:
Compare this request with the attached return policy. Identify the relevant rule, the facts supplied by the customer, and any missing information. Give a brief recommendation supported by the policy. If the evidence is incomplete, mark the case “Needs review.”
This produces an answer you can inspect. You can check the rule and see whether the facts support the recommendation.
An instruction to “think step by step” is not a guarantee of accuracy. A long explanation can still be wrong. Ask for the calculations, source passages, or short justification needed to verify the result.
For current information, use a tool that can consult current sources. For arithmetic, ask it to use an available calculator or code tool and state the inputs. If the necessary information or tool is unavailable, the response should say so.
3. Show what a good answer looks like
Words such as “friendly,” “professional,” and “concise” can mean different things to different people.
Examples make your preference easier to understand. Giving a few input-and-output examples is often called few-shot prompting.
For a plain-language rewrite, you might provide:
Original: “We invite interested individuals to participate in our upcoming garden maintenance initiative.”
Preferred: “Come help us get the garden ready for fall.”
Original: “Prior horticultural experience is not a prerequisite.”
Preferred: “You do not need gardening experience.”
Rewrite the following paragraph in the same clear, welcoming style. Preserve its facts.
The examples show what “plain language” means for this job.
Choose examples carefully. If each one removes important details, the tool may learn the wrong lesson. A short answer still needs to say enough.
4. Describe the shape of the answer
Think about where the result will go.
A comparison is often easier to use as a table. Meeting notes may need decisions, tasks, owners, and due dates. An email needs a subject line and a body someone can send.
Instead of “Summarize these notes,” try:
Create a table with four columns: task, owner, due date, and supporting note. Include only tasks agreed to in the supplied notes. Write “Not specified” when an owner or date is missing.
This reduces the work you must do after receiving the answer.
When software will read the output, developers may use a structured format such as JSON and supported schema controls. A schema defines the required fields and types of values. The application should still check the result before acting on it.
For everyday use, a short list or simple table is usually enough. Choose a format because it helps you use the answer.
5. Give missing information a sensible place to go
AI can produce a confident statement that is unsupported or false. This is often called a hallucination.
Clear prompts can reduce some errors. No wording can promise that every answer will be correct.
Give the tool an explicit way to handle gaps:
Use only the supplied event details. If the location is missing, put “[LOCATION NEEDED]” in the draft. Do not invent an address, a contact name, or a promise of refreshments.
That is more useful than simply saying “Do not make mistakes.” It defines what to do when the source material is incomplete.
You can also ask it to separate confirmed facts, assumptions, and open questions. Check names, dates, prices, quotations, and source links before using factual work.
Return to our garden invitation. A made-up promise of free lunch may sound harmless on the screen. It becomes a real problem when volunteers arrive expecting sandwiches.
6. Save the instructions you use repeatedly
If you write a weekly email, you should not need to explain your preferred style from scratch each time.
Keep the stable instructions together:
Write for local residents in clear, friendly language. Use short paragraphs. Keep the body under 150 words. Include one clear action the reader can take. Use only confirmed event details and flag missing information.
Then provide the details that change: this week’s event, date, location, and request.
Some tools offer project instructions or dedicated instruction settings. Developers can separate standing instructions from new inputs through the application’s supported message structure. In an ordinary chat, a saved template is a useful start.
Typing “SYSTEM” above a paragraph does not turn it into a privileged system instruction. Use the actual features your tool provides.
Keep templates short enough to understand. Remove old rules when they no longer serve the task.
A simple template you can reuse
Copy this and replace the bracketed text:
Task: [What should the AI produce or help decide?]
Audience: [Who will read or use the result?]
Context: [What background matters for this task?]
Source material: [Paste or attach the information it should use.]
Output: [State the format, length, and level of detail.]
Rules: Preserve the supplied facts. Identify missing information.
Separate assumptions from confirmed facts. Ask a question if a
missing detail would materially change the answer.
Success looks like: [Describe what makes the result useful.]
You do not need every line for every request. “Rewrite this paragraph in shorter sentences while preserving all the facts” may be all a small edit needs.
Improve the result with specific feedback
Your first prompt is a starting point.
If the response is too formal, identify what should change:
Replace the formal language with words I would use when talking to a neighbor. Keep the event details and the request to reply.
If it misses the purpose:
The main concern is that beginners may feel unqualified. Make it clear they can help without experience. Do not add new promises about training or equipment.
For a recurring task, save several realistic examples and compare results when you revise the prompt. Include incomplete and awkward inputs. A template that works for one tidy example may struggle with the next.
Anthropic’s prompting overview recommends defining success and testing against it before refining instructions. It also notes that some problems are better addressed by changing the model or system, rather than the prompt. Read the official prompting overview.
A clearer request cannot supply a missing document. An expert role cannot create access to your records. Choose a tool with the capabilities the job requires, then give it the information it needs.
Choose the right workspace for the request
An updated prompting habit is to consider what the assistant can actually access and produce.
If the task depends on a folder of reports, supply those sources or use a workspace designed to consult them. Gemini Notebook, formerly NotebookLM, is one option for working with a chosen source collection. Ask for source references and inspect them.
If you need a finished document or analysis across connected tools, ChatGPT Work and Claude Cowork provide capabilities beyond a plain text exchange. The available files, connections, and permissions still determine what a particular task can do.
The request can stay simple: “Use these three reports to prepare a two-page comparison. Link each important finding to its source and identify disagreements.” Give the system the reports, explain the result you need, and verify the output.
Choosing an appropriate workspace and providing the right material may help more than adding another page of instructions.
Picture that neighbor heading to the store again.
This time, you hand over a short list: carrots, onions, beans, and vegetable broth. You mention how many people are coming and ask them to call if something is unavailable.
That is the habit worth bringing to AI: explain the job, include what matters, and make uncertainty easy to flag.
You are much more likely to get the soup you had in mind.
Continue with autonomous workflows to see how repeat tasks can fit together, or browse the AI resource directory for tools to explore.
Put prompting to work in a business
Try writing and testing prompts, answering from sources, and evaluating output. These chapters develop business exercises; this introduction remains useful for everyday tasks.