How to Talk to AI (And Get Better Answers)

You type “write me a marketing email” into ChatGPT and get back something that reads like it was generated by a committee of robots who have never actually bought anything. It’s technically correct. It’s also completely useless. So you try again, maybe adding “make it better” — and somehow it gets worse.

Meanwhile, your colleague sends you a polished competitive analysis that she got from the exact same tool in one shot. What does she know that you don’t?

The difference isn’t intelligence or luck. It’s how she talks to the AI. The way you phrase a request — your “prompt” — is the single biggest factor in whether you get something brilliant or something you immediately delete. And it turns out that getting this right isn’t just about better results. It’s also about keeping your data safe.

The way you phrase a request is the single biggest factor in whether you get something brilliant or something you immediately delete.

How does AI actually read your words?

Before we get into techniques, a quick peek under the hood helps explain why wording matters so much.

When you type a prompt, the AI doesn’t read it the way you do. It breaks your text into chunks called “tokens” — roughly three-quarters of a word each. The word “unbelievable” becomes multiple tokens. Then, based on patterns learned from enormous amounts of text during training, the AI predicts the most likely next token, then the next, then the next. It’s essentially a text-prediction engine running at enormous scale.

This is why vague prompts produce vague answers. When you say “write something about AI,” there are millions of equally probable directions the model could go. It picks one at random. But when you say “write a 300-word summary of how small businesses use AI chatbots to reduce customer service costs,” you’ve narrowed the probability space to a much smaller, more useful region.

There’s also a practical limit to how much the AI can “hold in its head” at once — its context window. Depending on the model, this ranges from about 3,000 words to over 150,000 words. Go beyond that limit and the AI starts forgetting what you said at the beginning of the conversation. Think of it like talking to someone with a notepad: if the notepad fills up, earlier notes get pushed off the page.

What makes a good prompt?

The biggest mistake people make is being too casual. We talk to AI the way we’d talk to a coworker who already knows our project, our company, and our preferences. But AI doesn’t know any of that. Every conversation starts from scratch.

Three things consistently separate good prompts from bad ones: clarity, context, and format. Let’s look at each.

Be specific about what you want

Here’s a real before-and-after example:

Before: “Analyze our sales data.”

After: “Analyze Q3 2024 sales data to identify the top 3 performing product categories. Calculate month-over-month growth rates and provide 3 specific recommendations for improving underperforming segments. Present findings as a structured report with an executive summary, data analysis section, and action items.”

The first prompt could produce anything from a two-sentence summary to a rambling essay about sales theory. The second one tells the AI exactly what to analyze, what to calculate, how many recommendations to give, and how to structure the output. You’ll get something you can actually use.

The pattern is simple: specify the task (what to do), the scope (how much to cover), and the format (how to present it).

Give the AI context it doesn’t have

AI doesn’t know who you are, who your audience is, or why you’re asking. Providing that background dramatically changes the quality of what you get.

Before: “Write marketing copy.”

After: “You’re a digital marketing specialist for a B2B software company. Write an email targeting IT directors at mid-size companies (500-2,000 employees) who are evaluating cybersecurity solutions. Introduce our threat detection platform, emphasize ROI and ease of implementation, and include a call-to-action for a free security assessment. Tone: professional but approachable, no jargon.”

The first version gives the AI nothing to work with. The second tells it who it’s writing as, who it’s writing for, what to emphasize, and how to sound. The output will be something you could actually send to a prospect.

Think of context as having three layers: role (who the AI should act as), situation (the background and constraints), and audience (who will read the output).

Tell it how to format the response

Nothing is more frustrating than getting useful information buried in a format you can’t use. If you need bullet points, say so. If you want a table, ask for one. If you need it under 200 words, specify that.

Good format instructions look like: “Present as a numbered list,” “Structure as a two-column comparison table,” “Limit to 3 paragraphs,” or “Format as an email with subject line.”

Specify the task, the scope, and the format. Those three things separate prompts that produce usable output from prompts that waste your time.

Techniques that make a real difference

Once you’ve got the basics of clarity, context, and format, there are a few power moves that can take your results to another level.

Ask it to think step by step

This one sounds almost too simple, but it works remarkably well for complex problems. When you add “let’s work through this step by step” to a prompt, the AI produces noticeably more accurate and thorough responses. Research suggests this “chain-of-thought” approach can improve accuracy on reasoning tasks by 40-60%.

Why? Because it forces the AI to show its work rather than jumping to a conclusion. Each step constrains the next step, reducing the chance of a wrong turn.

Try it with something like: “I need to decide whether to expand into the European market. Work through this step by step: first, analyze market size and growth potential. Second, evaluate competitive advantages and challenges. Third, assess required investment and timeline. Fourth, calculate potential ROI. Finally, give a clear recommendation with reasoning.”

Show it what you want (few-shot examples)

Sometimes it’s easier to show than tell. Instead of describing the exact format or style you want, you can provide an example or two and say “do it like this.”

This technique — called “few-shot learning” — is especially powerful when you need a consistent style that’s hard to describe in words. Give the AI 2-3 examples of the kind of output you want, then ask it to produce more in the same pattern.

For instance, if you need customer service responses in your company’s specific voice, show the AI one or two real responses you’ve written and then ask it to handle a new scenario the same way. Two good examples usually outperform a paragraph of abstract instructions.

Refine instead of starting over

Your first prompt rarely produces a perfect result, and that’s fine. The real skill is knowing how to refine rather than starting from scratch.

When an AI response is 70% right, don’t throw it away and rewrite your prompt. Instead, tell the AI specifically what to fix: “The data analysis section is strong, but the recommendations are too vague. Rewrite the recommendations section with 3 specific strategies we could implement in the next 90 days. For each one, include the expected timeline, resources needed, and likely impact.”

This iterative approach works because each round of feedback builds on what’s already working. Think of it like editing a draft with a coworker rather than asking them to start over every time.

Chain-of-thought prompting can improve accuracy on reasoning tasks by 40-60 percent. Adding five words to your prompt makes the difference.

Keep your data safe while prompting

Here’s the part most prompting guides skip entirely: security. As we covered in Your Data and AI, nearly half of all prompts sent to AI tools contain sensitive information — customer names, employee data, financial details. People paste in things they would never email to a stranger, without thinking twice.

Watch what you share

Before you hit enter, scan your prompt for anything you wouldn’t want posted on a bulletin board. Real names, account numbers, proprietary data, medical details — all of it should either be removed or replaced with placeholders.

Instead of “analyze John Smith’s account #4829-3847 showing revenue of €2.1M,” try “analyze a customer account showing annual revenue of approximately €1.8M.” You get the same quality of analysis without exposing real data.

Don’t mix sensitive and untrusted content

This one matters more than most people realize. If you paste a document from an unknown source into the same conversation where you’ve been discussing confidential business strategy, you’re creating a risk. That document could contain hidden instructions — a technique called “prompt injection” — designed to manipulate the AI into revealing or misusing the information you’ve already shared.

Prompt injection works because AI treats everything in the conversation as instructions. A cleverly worded paragraph buried in a document can override the AI’s original behavior. Researchers have demonstrated attacks where injected instructions caused AI to extract confidential information, generate malicious code, or take unauthorized actions through connected tools.

The defense is straightforward: keep separate sessions for separate sensitivity levels. Don’t process a random PDF from the internet in the same chat where you’re analyzing your company’s financial data.

Verify what comes back

As we explored in When AI Gets It Wrong, AI fabricates information more often than most people realize. Always check AI outputs before acting on them, especially if they contain any of the following red flags: requests for additional information you didn’t plan to share, suggestions to contact unfamiliar parties, recommendations that seem outside the scope of what you asked, or responses that feel oddly formatted or off-topic. Any of these could indicate the AI’s behavior was manipulated — or simply that it generated something unreliable.

What about temperature and settings?

If you’ve poked around in AI settings, you may have seen something called “temperature.” This controls how creative versus predictable the AI’s responses are. Think of it as a dial between “strict librarian” and “jazz musician.”

Low temperature (0.1-0.3) produces consistent, predictable responses — great for factual work, technical documentation, or anything where accuracy matters more than flair. High temperature (0.8+) produces more creative, varied, and sometimes surprising outputs — better for brainstorming, creative writing, or generating fresh ideas. Most general-purpose conversations work well in the middle range (0.4-0.7).

You don’t need to obsess over this setting, but knowing it exists helps explain why the same prompt sometimes gives you different results, and why turning down the temperature can help when you need reliability over novelty.

Build a system, not a habit

The people who get the most out of AI don’t craft perfect prompts from scratch every time. They build templates for the tasks they do repeatedly.

If you write weekly status reports, create a prompt template with slots for this week’s accomplishments, next week’s priorities, and blockers. If you regularly analyze data, build a template that specifies the analytical framework, output format, and evaluation criteria. If you draft client emails, set up a template that includes your brand voice guidelines and typical structure.

Templates save time and — more importantly — they encode best practices so you don’t have to remember them. They include the security checks, the format specifications, and the context that makes AI responses actually useful.

What to do now

1. Pick one task you already use AI for and rewrite the prompt using the clarity-context-format framework. Compare the results with what you were getting before.

2. Try chain-of-thought prompting on your next complex question. Add “work through this step by step” and see how the response quality changes.

3. Audit your last 5 AI prompts for sensitive data. Did you share names, numbers, or proprietary information that could have been anonymized? Practice swapping in placeholders.

4. Create one reusable prompt template for a task you perform weekly. Include the role, context, format specification, and a reminder to sanitize sensitive data before filling it in.

5. Start keeping a “prompt journal” — even a simple notes file — where you save prompts that worked well. Over time, this becomes your personal playbook for getting better results faster.

The goal here isn’t to become a “prompt engineer.” It’s to communicate clearly enough that AI actually helps you instead of wasting your time — and to do it without accidentally handing over information you’d rather keep private. Those two skills, it turns out, go hand in hand.