What a prompt is — how you "tell" AI changes the result
When you ask a friend for a favor, saying "just do something nice" gets a very different result than saying "I'd like a two-page A4 proposal by next Friday, with a budget under $5,000," doesn't it?
Instructions to AI are the same. A prompt is the instruction you send to AI. And a single choice in how you write the prompt can change the AI's answer dramatically.
This isn't just a feeling. A meta-analysis reviewing over 1,500 research papers on prompting found that simply adding specific conditions to a prompt raised answer accuracy by about 30% and reduced irrelevant information by 42%[1].
You may have heard the term "prompt engineering." There's even data showing that in 2025, 68% of companies incorporated prompting skills into standard training for all employees[2]. It's no longer a specialized skill for engineers alone — it has become a basic skill every AI user needs.
In this chapter, we explain how to write prompts that anyone can put into practice today, with concrete examples. For an even more practical set of templates, see our practical guide to prompt engineering as well.
The prompt transforms the AI's answer
The 5 elements of a good prompt
Effective prompts have five elements. In a cooking-recipe analogy, they're like "who makes it (role)," "what to make (task)," "who it's for (context)," "how to plate it (format)," and "allergy accommodations (constraints)."
Just prefacing with "You are an expert in X" changes the AI's tone and knowledge level.
Example: "You are a web marketer with 10 years of experience. Tell me how to grow the page views of a personal blog."
Not just "tell me" or "think about it" — make the verb concrete. "Summarize," "compare," "propose five" — the clearer the action, the more precisely the AI acts.
Telling the AI your situation greatly improves the direction of the answer. Just writing "I'm a university student with no programming experience" automatically adjusts the difficulty.
"In bullet points," "as a table," "within 300 words," "with headings" — without a specified format, AI tends to reply in long paragraphs. Specify to match your use.
Passing along NG rules and conditions together — "don't use jargon," "limit it to the domestic market," "always include the downsides" — reduces do-overs.
💡 You don't need all five every time. While you're getting used to it, just being mindful of the two — ② task + ③ context — already changes answer quality a lot. Add the other elements as you get comfortable.
A practical toolkit — 7 patterns you can use today
A large-scale 2024 survey, "The Prompt Report," systematically classified as many as 58 prompting techniques[3]. But you don't need to memorize 58. Only about seven are truly useful for everyday AI work. Here they are, organized by difficulty.
Add numbers, targets, conditions. The simplest, with the biggest effect.
Bullets, tables, word count. Receive it in an easy-to-use form.
"You are an X with Y years of experience" changes the depth of the answer.
2–3 input/output examples work well for classification, conversion, and routine tasks.
Boosts reasoning accuracy on calculation, logic, and complex analysis.
Have it critically re-check its answer to raise accuracy.
Have the AI write the prompt itself, then tune it for your needs.
[Beginner] ① Give specific instructions
The simplest technique, with the biggest effect. Just add numbers, targets, and conditions to a vague question.
❌ Vague: "Tell me how to increase sales"
✅ Specific: "Give me 5 measures to raise the monthly sales of a 10-employee ramen shop by 20%, ordered from lowest cost. The current average spend per customer is about $6, and monthly footfall is about 3,000."
[Beginner] ② Specify the output format
Research shows that the presence or absence of a format specification can change accuracy by as much as 76 points in some cases[1]. Just adding "in bullet points," "as a comparison table," or "within X words" makes the answer easy to use.
[Beginner] ③ Assign a role (role-play)
Just telling it "you are X" gets you an answer specialized to that field.
"You are a sales manager with 10 years of experience. Tell me, in order of priority, the skills a new salesperson should acquire in their first three months."
The key is to include years of experience and a specific position. "You are X with Y years of experience" changes the depth of the answer more than just "you are an expert."
[Intermediate] ④ Show examples (few-shot)
A method where you show the AI two or three input/output examples before asking the real question. It's especially powerful for classification, conversion, and routine tasks.
Classify the sentiment of product reviews in the following format.
Review: "This product is the best!" → Sentiment: Positive
Review: "It was disappointing." → Sentiment: Negative
Review: "It works fine, I guess." → Sentiment: NeutralReview: "Delivery was fast, but there was a scratch." → Sentiment:
There's an interesting research finding: the effect of few-shot depends more on the "diversity" of examples than their "correctness." In other words, showing the breadth of the pattern matters more than preparing perfect examples[4].
[Intermediate] ⑤ Make it think step by step (Chain-of-Thought)
Research shows that just adding "think step by step" raises accuracy on reasoning tasks by about 35%[1]. It's especially effective for calculation problems, logic problems, and complex analysis.
"An apple is $1.50 each and an orange is $0.80 each. If I buy 3 apples and 5 oranges and pay with a $10 bill, how much change do I get? Answer while showing the calculation one step at a time."
⚠️ A 2025 caveat: For the latest "reasoning models" — such as ChatGPT's o-series and Claude's Extended Thinking — do not instruct them to "think step by step." These models perform step-by-step thinking automatically and internally, so explicit instructions can actually hurt performance[5]. Use "think step by step" with standard conversational models that lack a reasoning mode.
[Advanced] ⑥ Have it self-review
A technique where you have the AI critically check its own answer. It's effective in situations that demand accuracy.
(After receiving the AI's answer)
"Critically check this answer for problems and omissions. If there are errors, correct them."
That said, the AI can't find 100% of its own mistakes. Researchers point out a "plausibility trap": the more plausible the AI's output looks, the harder it is to verify[5]. For important judgments, always include a human check too.
[Advanced] ⑦ Meta-prompting — have the AI write the prompt
Once you're comfortable writing prompts, there's also a method of having the AI write the prompt itself.
"I'm a freelance web designer, and I want to efficiently create proposals for clients. Create the optimal prompt to achieve this goal."
The trick is not to use the AI-written prompt as-is, but to tune it to your own situation.
Before / After — the same question, transformed
Since technique descriptions alone are hard to feel, let's look at concrete Before/After examples. On the left is a vague instruction; on the right is an instruction with conditions added.
Example 1: Writing a business email
| Before | After |
|---|---|
| "Write an apology email" → A generic, unusable email is generated |
"Write an apology email to Mr. Tanaka, a department head at a client, about a delivery being delayed by three days. The cause was a logistics issue, and it's now expected to arrive next Monday." → An email you can send right away is done |
Example 2: Brainstorming a project
| Before | After |
|---|---|
| "Come up with new product ideas" → Unclear which industry or product, so the suggestions are vague |
"Propose 5 subscription health-food products under $30/month aimed at women in their 20s–30s. Also add the points that differentiate them from competitors A and B." → Ideas you can drop straight into a proposal |
Example 3: Learning and research
| Before | After |
|---|---|
| "Teach me machine learning" → A stiff, hard-to-follow explanation comes back |
"Explain how machine learning works so that a middle schooler can understand, using a cooking-recipe analogy. Don't use jargon, and include two concrete examples." → An approachable explanation you can grasp right away |
The common point is making "to whom, what, how, and in what form" clear. The more conditions you add, the more precise the AI's answer.
Common failure patterns and how to fix them
Here are five traps people fall into when they start using AI, and how to get out of them.
❌ Failure 1: Instructions that are too vague
"Write something nice" or "make it better" are instructions that trouble the AI too. Convey specifically what makes it "good," and how. "In a casual tone," "include three numbers," "within 300 words" — the trick is to make your criteria explicit.
❌ Failure 2: Asking everything at once
"Tell me how to build a marketing strategy, recommended tools, and how to set a budget" — with this, the AI tries to cover everything shallowly, and none of it lands. Ask about one theme per question. Building the next question on the previous answer yields deeper responses.
❌ Failure 3: Believing the AI's output as-is
AI can be wrong with full confidence (hallucination). Especially watch out for:
- Specific numbers and statistics — ask for and confirm the source
- Proper nouns — always verify people's names, company names, and law names
- The latest information — the model's knowledge has a cutoff
- Law and taxation — always confirm with official information
The countermeasure is simple. Get into the habit of asking "tell me the source of this information." Information the AI can't cite a source for is a sign to verify.
❌ Failure 4: Ending after a single exchange
"I asked AI but the answer was mediocre" → "See, AI is useless." This is a huge waste. As we'll explain in detail in the next section, dialogue with AI assumes two or three rounds.
❌ Failure 5: Dumping everything on the AI
What AI is good at: drafting, organizing information, and generating ideas. What humans are good at: final judgment, reading context, and ethical consideration. Combining the two is the most effective way to use it. "Have AI make the draft and let a human finish it" — this division of labor works best.
The art of dialogue with AI — don't stop at one shot
More important than prompting technique is the mindset of "layering the dialogue."
Prompting research finds that iterative feedback raises output quality by 35%[1]. Rather than demanding perfection on the first try, look at the AI's answer, correct the direction, and ask again. This cycle raises quality.
Send the first prompt, mindful of the 5 elements.
Judge the good points and what's missing with human eyes.
Convey course corrections, deeper dives, and have it revise.
Reach the quality you're after in 2–3 rounds.
6 types of feedback
For those who "don't know what to give feedback on," here are six easy-to-use types.
| Type | Example of use |
|---|---|
| Course correction | "Make the style more casual" / "This is too technical; make it beginner-friendly" |
| Deeper dive | "Explain the third idea in more detail" / "Add concrete steps" |
| Added condition | "Narrow it to ideas under a $10,000 budget" / "Limit it to the domestic market" |
| Change of perspective | "Reconsider from the customer's point of view" / "Also give the opposing view" |
| Quality check | "Check this content for errors or contradictions" |
| Format change | "Turn this content into an email" / "Reorganize it into a table" |
With two or three rounds of feedback, answer quality improves dramatically. The trick is not to demand perfection from the start.
From prompting to "context engineering"
Finally, let's touch on the latest trend in prompting technique.
In June 2025, AI researcher Andrej Karpathy said that "everyday short prompts are only a tiny part of industrial-scale AI use," and proposed the concept of context engineering[6].
This is the idea of properly designing not just the prompt (instruction) but the entire body of information you give the AI — reference materials, past exchanges, tool outputs, and so on. For example, having the AI read internal documents before asking a question, or attaching past meeting notes before requesting a summary. For details, see What is context engineering. This idea also becomes important in vibe coding, where you leave the implementation to AI, and in spec-driven development.
Still, the basics don't change. First, put into practice the five elements and seven techniques you learned in this chapter. That alone should transform the quality of the AI's answers beyond recognition.
- A single choice in how you write the prompt transforms the answer. Just adding conditions raises accuracy.
- A good prompt has five elements: role, task, context, format, constraints. Start with ② task + ③ context.
- The practical techniques come in seven, by difficulty. From specifics, format, and role-setting to few-shot, step-by-step thinking, and self-review.
- For the latest reasoning models, "step by step" is unnecessary. Use it on standard models.
- Don't stop at one shot; polish with 2–3 rounds of feedback. Beyond that lies context engineering.
References
- Gupta, Aakash. "I Spent a Month Reading 1,500+ Research Papers on Prompt Engineering." Medium, 2025.
- "Is Prompt Engineering Dead?" Fast Company, May 2025.
- Schulhoff, Sander et al. "The Prompt Report: A Systematic Survey of Prompting Techniques." arXiv:2406.06608, 2024.
- Min, Sewon et al. "Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?" arXiv:2202.12837, 2022.
- Lakera. "Prompt Engineering Guide 2026." lakera.ai, 2026.
- Karpathy, Andrej. "Context Engineering." X (Twitter), June 2025.
Related links:
- Prompt Engineering Guide — a comprehensive guide to prompting techniques (English)
- Anthropic Prompt Engineering Docs — a prompt-design guide for Claude
In the next chapter, we'll introduce how to put the prompting techniques from this chapter to work in real situations — concrete uses at work, in learning, and in creative work.