Silver fountain pen nib drawing one glowing hot pink line across black paper, a clear single instruction like well-structured AI prompts, Amelia S. Gagne
Ai Getting Started • Updated • 16 min read

How to Write AI Prompts That Give You Useful Answers

Structured AI prompts scored about 4.8 out of 5 for usefulness in a 2026 study of 105 business professionals, against 3.27 for plain requests. Here is the five-part structure, real examples, and the mistakes to skip.

Good AI prompts give the model what a new colleague would need: who it's working for, the background, one clear task, the limits and the shape of the answer. In a 2026 peer-reviewed experiment, 105 business professionals rated answers from plain one-line requests at 3.27 out of 5 for usefulness, and answers from structured AI prompts at about 4.8.

Silver fountain pen nib drawing one glowing hot pink line across black paper, a clear single instruction like well-structured AI prompts, Amelia S. Gagne
The model in that 2026 study stayed the same for every rating. Only the AI prompts changed, and usefulness rose from 3.27 to 4.82 out of 5.

Most people blame the AI when they get a bad answer. Usually the problem is the request. That's good news, because the gap between "this tool is useless" and "this just saved me an hour" is a skill you can learn, and it's the same skill that makes someone good at running a meeting or briefing a contractor.

Why do AI prompts get vague answers?

A model starts every conversation knowing nothing about your company, your customers or what you tried last week. When the request is short, it fills the gaps with the most average answer it can produce. That's where the generic, could-be-anyone output comes from.

The study above, by Alessia Cantini and Andrea De Mauro in Big Data and Cognitive Computing (July 2026), tested this on analysis and summary requests in Microsoft Copilot. Adding one example lifted the rating. Asking for step-by-step reasoning lifted it about the same. Doing both lifted it the most.

Bar chart of usefulness ratings for AI prompts by style, from 3.27 for a plain request to 4.82 for one example plus step-by-step reasoning, Amelia S. Gagne
Structured AI prompts scored higher than plain requests in every format tested (Cantini and De Mauro, 2026, 420 ratings from 105 business professionals).
How the prompt was writtenAverage usefulness (1 to 5)
Plain request, no example3.27
One example4.01
Several examples4.09
Step-by-step reasoning4.11
Several examples plus step-by-step reasoning4.79
One example plus step-by-step reasoning4.82

The study is small and most of its raters were in Italy, so treat the exact numbers with care. The direction matches what I see every day: the answer gets better when the request does.

The five-part structure for AI prompts

I've used AI daily in client work for years, from enterprise platforms to a small shop's launch plan. The structure I come back to has five parts, in this order:

  1. Role. Who should it be for this job? "You are a senior financial analyst" gets different work than "You are a marketing copywriter," even with the same data.
  2. Context. The background it doesn't have: the situation, the audience, what happened before, and the facts it must use.
  3. Task. Exactly what you want done. One verb, one deliverable.
  4. Constraints. Length, tone, what to avoid, what matters most.
  5. Format. What the answer should look like: a table, a draft email, a numbered list with headers. An example of a good answer is the strongest format instruction you can give.

A quick question might need two of the five. When an answer comes back vague or off target, though, the missing piece is almost always one of them. The vendors' own guides agree on the shape: Anthropic, OpenAI and Google each recommend a role, background, clear instructions and examples of the output you want.

What I add for marketing work

For anything a customer will see, I do the research before I write any of the AI prompts. That includes sentiment: how the audience actually talks about the product, the problem and the competition. Then the prompts carry the brand colors, the voice, and the company facts the piece depends on. Leave those out and you get something polished that could belong to any company.

Should you let AI interview you before it writes?

Yes, for anything that matters. Before I create a campaign, a proposal or a strategy document, I ask the AI to interview me first. It asks what it needs to know for that task, I answer, and only then does it draft.

The interview catches what I would have forgotten to say. When you write AI prompts, you know your business so well that half the important context feels too obvious to mention. Questions pull it out.

I need a launch email for a new service. Before you write anything, interview me. Ask me one question at a time about the audience, the offer, what makes it different, the tone, and anything else you'd need to write this well. When you have enough, summarize what you heard in five bullets and wait for me to confirm before drafting.

Asking for one question at a time keeps the answers thoughtful. The summary at the end is your chance to correct a misunderstanding before it ends up in a draft.

Give it your source of truth

Context is the part of AI prompts people skip most, and it's the part that makes the answer yours. I keep a few short documents for exactly this:

  • A source of truth document: company basics, contact details, branding, mission. The facts that should never come out wrong.
  • A voice document: the phrases I use and the ones I never would, how I structure an argument, what I like and dislike, and how I want information delivered to me (format, level of detail, how sources are cited).
  • Task documents: smaller references for recurring jobs, so the same standards apply every time.

Inside our company, I rarely paste any of it anymore. LTFI, the operating system we built at Kief Studio, holds that knowledge so every AI tool we use starts with it. When I'm on another device for a quick task, the documents come with me. If you use AI for the same kind of work again and again, keep a context block you can paste in each time. Your business already writes this down for new hires; institutional memory that survives turnover works for AI too.

Where the context goes matters too. In a 2024 study from Stanford and UC Berkeley, a model found an answer far more reliably when the document holding it sat at the start or end of a long prompt than when it was buried in the middle.

Line chart showing answer accuracy by position of the key document among 20 in a long prompt, highest at the start and end and lowest in the middle
Long AI prompts lose track of the middle: accuracy fell from 75.8 percent with the key document first to 53.8 percent with it tenth of 20 (Liu et al., Transactions of the ACL, 2024).
Position of the document with the answer (of 20)Accuracy (%)
1st75.8
5th57.2
10th53.8
15th55.4
20th63.2
No documents at all56.1

That test used a 2023 model, and newer ones handle long material better. The habit still holds: put the long material first, and the facts that matter and the actual request where they can't be missed.

Before and after: AI prompts that work

These are built on the kinds of tasks teams hand to AI every week.

A client follow-up email

Before:

Write a follow-up email to a client.

After:

You are a project manager at a technology consulting firm. We just held a kickoff with a new client, a fintech company with 80 employees, about moving their old reporting system to a modern data pipeline. Write a follow-up email that summarizes the three priorities we agreed on (data audit, stakeholder interviews, tool evaluation), confirms our next meeting for Thursday at 2 PM, and asks who on their team should be invited to the data audit session. Tone: professional and warm. Under 200 words.

That takes about 90 seconds to write, and the email needs light edits instead of a rewrite. Notice what it doesn't ask for: passwords or system logins. Those never go through an AI tool or an email draft.

Marketing for a new product

Before:

Create a brand name, target audience, marketing strategy, and 10 Instagram captions for my new candle business.

I used to write AI prompts like this one. The answers covered everything and said nothing: a forgettable name, an audience of "people who like candles," and captions that could have belonged to any shop. Four jobs in one prompt meant each got a fraction of the attention.

After, as a sequence:

Step 1: You are a brand strategist. Here is my research on the candle market in my area, including what customers say they love and complain about in reviews. Interview me about my product, my price point and why I started, one question at a time. Then describe the three customer groups most likely to buy, with what each one cares about.

Step 2 (new message, after I pick a group): Using the customer group I chose and my voice document below, suggest ten brand names with a one-line reason for each. Avoid anything that sounds like the three competitors I listed.

Step 3: Using the name I chose, my brand colors (deep green and cream) and my voice document, write five Instagram captions for the launch week. Each one under 150 characters, one call to action each, no hashtags in the first line.

Each step feeds the next, and each answer is better because each of the AI prompts has one job.

Research with sources

The first prompt that made me think "this changes everything" came a couple of years before ChatGPT. I was writing for a blog where accuracy mattered, and I asked an AI tool for sources on my topic. It handed back a collection of citations in seconds. Hours of searching became minutes, and I didn't have to click through unfamiliar sites to find them.

Today I'd add one rule to AI prompts for research: open every source before you use it. Models can produce citations that look perfect and point to nothing, and no provider is immune. In a peer-reviewed test of eight free AI chatbots, published in the Journal of Data and Information Science in 2026, Álvaro Cabezas-Clavijo and Pavel Sidorenko-Bautista checked 400 references by hand. Overall, 39.8 percent were wrong or made up.

Horizontal bar chart of wrong or fabricated references from eight AI chatbots tested in February 2025, from 100 percent for Microsoft Copilot to 0 percent for Grok and DeepSeek, a reason AI prompts for research should demand sources you can open
Across eight chatbots, 39.8 percent of AI-supplied references were wrong or fabricated, and results ranged from 0 to 100 percent (Cabezas-Clavijo and Sidorenko-Bautista, 2026). Write AI prompts for research that ask for sources you can check.
Chatbot (model tested, February 2025)References wrong or fabricated (%)
Microsoft Copilot (free, GPT-4)100
Perplexity (Sonar)72
Claude (3.5 Sonnet)64
ChatGPT (GPT-4o mini)38
Le Chat (Mistral Large)32
Gemini (2.0 Flash)12
Grok (Grok-3)0
DeepSeek (V3)0
All eight together39.8

Each chatbot was tested on just 50 references using its free version, and these models have since been replaced. Rankings also change from study to study, so no provider gets a pass.

Find primary sources on how small businesses use AI for customer service, published in the last two years. For each one, give the title, the publisher, the date, a working link, and the exact sentence the figure comes from. If you can't find a primary source for a claim, say so instead of guessing. I will open every link before I use it.

AI prompts that ask for the exact sentence make checking fast. Qurtoo's guide to why AI makes things up has three more checks worth running on anything you'll publish.

Five mistakes that make AI prompts fail

1. Being too vague

"Help me with my marketing" could produce a brand strategy or a single tweet. Narrow it: "Draft three subject lines for a product launch email to CFOs at companies with 200 to 500 employees." Small wording choices matter more than most people expect. In the 2022 NeurIPS paper that introduced "Let's think step by step," changing one line before the answer moved accuracy on the same 600 math problems from 9.3 percent to 78.7 percent.

Horizontal bar chart comparing accuracy when different single lines were added to AI prompts, helpful instructions near 75 percent and irrelevant or misleading lines below 20 percent, Amelia S. Gagne
One line in AI prompts can make or break the answer: helpful wording reached 78.7 percent accuracy while misleading wording fell to 9.3 percent (Kojima et al., NeurIPS 2022).
Line added to the promptTypeAccuracy (%)
Let's think step by step.Helpful78.7
First,Helpful77.3
Let's think about this logically.Helpful74.5
Let's thinkHelpful57.5
Don't think. Just feel.Misleading18.8
(No added line)Baseline17.7
By the way, I found a good restaurant nearby.Irrelevant17.5
It's a beautiful day.Irrelevant13.1
By using the fact that the earth is round,Misleading9.3

That paper tested a 2022 model, and today's models often reason step by step without being told. The lesson that lasts is how much one word can steer the result. When a workflow is hard to put into words, I draw it: a quick diagram or sketch, attached to the prompt, describes a process more precisely than three paragraphs, and it saves a lot of time correcting an answer that went off in the wrong direction.

2. Skipping context

Without your industry, audience, constraints and recent decisions, the answer will be generic at best and misleading at worst. This is what the source of truth and voice documents above are for. Thirty seconds of setup saves five minutes of revisions.

3. Asking for too many things at once

The candle prompt above is the classic case. One of those AI prompts asking for a strategy, a calendar, three emails and a blog post gets shallow work on all four. Do the strategy first, then hand that output to the next prompt. If you want to go deeper on scoping work this way, the same principles apply to context switching in general.

4. Borrowing someone else's voice

This is the mistake I see most when businesses start using AI for content. They paste in a brand they admire and ask for "something like this." That brand has a different audience, a different product, different prices, a different mission and a different culture, and all of it shows up in the writing. Copy it and your website and social posts end up talking to someone else's customers.

Give the AI your own examples, your voice document and your customer research instead. It's the same principle behind using AI for customer communication without sounding robotic: the draft should sound like the company that's sending it.

5. Not iterating

The first answer is a draft. My follow-ups are rarely "make it better." I ask the AI to find the holes: everything I could have missed, the questions a skeptical reader would ask, and the points of view that disagree with mine. Then I ask for an unbiased version. Everyone has opinions, and a piece is only complete once it accounts for the other perspectives and the related topics around it.

Review this draft as a skeptical expert in the field. List every gap, unsupported claim and missing perspective, including views that disagree with mine. Then list related topics a reader would expect me to address. Don't rewrite yet.

Specific feedback pays off. In the Self-Refine study (NeurIPS 2023), models that critiqued and revised their own drafts improved with every round, and the first round brought the biggest jump.

Line chart of a writing task score rising from 29.0 percent on the first draft to 49.7 percent after three rounds of feedback, showing why AI prompts should include follow-up rounds
Follow-up AI prompts pay off: coverage of required concepts rose from 29.0 percent on the first draft to 49.7 percent after three rounds of feedback (Madaan et al., Self-Refine, NeurIPS 2023).
RoundRequired concepts covered (%)
First draft29.0
After round 140.3
After round 246.7
After round 349.7

I also use more than one AI. Each model is strongest at different work, so I match the tool to the task, which is why our company system works with whichever AI you choose, including one you run yourself.

What should never go into AI prompts?

The things I trust AI with least are the things that have to stay private: passwords and secrets, private emails and messages, financials, and intellectual property. Leave those out of AI prompts unless the tool has been set up and approved for that kind of data, with access scoped to the job.

Even when a provider says it doesn't keep your data, I still take the precautions. A policy statement doesn't replace knowing where the data goes and who can reach it. That care is why we built AI governance into everything at Kief Studio, for our own work and for clients. If you're setting this up, start with what AI can and can't do with your business data, then write it down in an AI policy your team will follow. The AI implementation work at Kief Studio starts with the same question: what should this tool be allowed to touch?

Prompting is a management skill

I've had management responsibilities for most of my career, and it made prompting feel familiar from the start. Good AI prompts are good delegation: the person or the tool knows the goal, the context, the standard and the deadline, and knows when to come back with questions.

Most workers haven't built the habit of writing AI prompts yet. When Pew Research Center surveyed 5,273 employed US adults in October 2024, 55 percent said they rarely or never use AI chatbots at work.

Donut chart of how often US workers used AI chatbots at work in 2024, 9 percent daily or weekly, 7 percent monthly, 55 percent rarely or never and 29 percent unaware
55 percent of US workers rarely or never used AI chatbots at work in late 2024 (Pew Research Center, 5,273 adults), so a team that writes good AI prompts is still ahead.
How often workers use AI chatbots at workShare of employed US adults (%)
Every day or a few times a week9
A few times a month7
Rarely or never55
Haven't heard of AI chatbot use at work29

Experience alone won't close that gap. In the Cantini and De Mauro study, how often people used AI made no measurable difference to how much structure helped. The authors conclude that training and shared standards for AI prompts may matter more than learning by trial and error, which is the case for training your team on AI yourself.

If you're just starting, figuring out where AI fits is a good first step, and knowing what to try first and what to skip will save you some frustration.

AI won't replace knowing your business. It multiplies what you already bring to it: clear thinking, careful governance and the habit of delegating well.

Related reading

Frequently asked questions

How do I create an AI prompt?

Give the model a role, the background it needs, one clear task, your constraints and the format you want, ideally with an example of a good answer. For important work, let the AI interview you first, then review the draft and ask it to find what's missing.

Do I need a different prompt format for each AI tool?

No. The role, context, task, constraints and format structure works across ChatGPT, Claude, Gemini and most other models, and each vendor's own guide recommends the same parts. Tools differ in small ways, so if AI prompts work well in one, expect it to work reasonably well in another.

How long should AI prompts be?

Long enough to be specific and short enough to stay on one task. Rewriting an email needs two or three sentences. Analyzing data or drafting a strategy usually needs a paragraph of context plus clear constraints, with long reference material placed first and the request at the end.

Is prompt engineering a job I should hire for?

Usually not as a standalone role. The people who write the best AI prompts communicate clearly and know the business, which is the same skill set that makes them good at their current jobs. Training your team and agreeing on shared prompt standards tends to pay off faster.

What is an example of a good AI prompt?

"You are a project manager at a technology consulting firm. Write a follow-up email to a new fintech client that summarizes the three priorities we agreed on, confirms our next meeting for Thursday at 2 PM, and asks who should join the data audit session. Professional and warm, under 200 words." It names a role, gives the background, asks for one thing, sets limits and describes the answer.

What are examples of good AI prompts for business?

The before and after examples above cover a client follow-up email, a product launch done as a three-step sequence, and research that asks for sources you can check. Each one names a role, gives the background, asks for one thing, sets limits and describes the answer it wants.

How do I get better answers from AI prompts I already use?

Add the context you've been leaving out, split any prompt that asks for several things into steps, and follow every first answer with a request to find the gaps. If the same AI prompts come up every week, save them with your context block so the standard stays the same.

How do I write AI prompts for branding?

Start with research on how your customers talk about the product and the problem, then put your brand colors, your voice document and your company facts into the prompt. Use your own examples rather than a brand you admire, because their audience, prices and mission are different from yours and the writing will show it.

How do I get accurate answers from ChatGPT?

Give it the facts it should use instead of letting it guess, ask for a working link and the exact sentence behind every figure, and tell it to say so when it can't find a source. Then open the sources yourself, whichever AI you use. In a 2025 study by the European Broadcasting Union and the BBC, journalists reviewed about 2,700 answers to news questions and found a significant problem in 45 percent of them: 36 percent for ChatGPT, 37 percent for Copilot, 30 percent for Perplexity and 76 percent for Gemini.

Can AI help me write my prompts?

Yes, and it's one of the most useful habits to build. Describe what you're trying to get done and ask the AI to interview you, one question at a time, about everything it needs. Have it summarize what it heard before it drafts anything, so you can correct it early.

Can AI prompts include pictures or diagrams?

Yes. Most current AI tools accept images, so a quick sketch of a workflow or a chart can sit in the prompt next to the text. For a process that's hard to describe in words, a diagram is often the clearest way to show what you mean.

What's the biggest mistake beginners make with AI prompts?

Treating the AI like a search box. A short, keyword-style request gets an average answer. A request written like a brief to a capable colleague, with context and a clear goal, gets something you can use.

AI Getting Started Aug 19, 2026 • 8 min

Why AI Pilots Stall Before Production

AI pilots stall before production for boring reasons: no owner, no definition of done, no data path, no kill switch. CISA 1 May 2026: start low-risk. NIST AI RMF: Govern, Map, Measure, Manage.

Work With Us

Need help building this into your operations?

Kief Studio builds, protects, automates, and supports full-stack systems for businesses up to $50M ARR.

Newsletter

New writing, straight to your inbox.

Strategy, psychology, AI adoption, and the patterns that actually compound. No spam, easy to leave.

Subscribe