Prompt Engineering: The Practical Compendium — 6 Parts and Techniques to Get the Answers You Want from AI

You ask the same AI the same thing — yet one person calls it "useless" while another is amazed at how "almost too capable" it is. The real cause of that gap is often not the AI's power, but how the prompt (the instruction) is written. AI is like a brilliant new hire who cannot read your mind: the quality of the instruction you give largely determines the quality of the answer you get back.

This article is a practical compendium of that skill — prompt engineering — organized so a beginner can use it right away. In short, the trick is to keep "the 6 parts of a good prompt (role, context, instruction, examples, format, constraints)" in mind, and to refine through conversation rather than nailing it in one shot. We boil down the approach that the public guides from OpenAI, Anthropic, and Google all recommend, in plain language, onto a single page. For concrete cases, see also prompt tips for asking AI to build an app, and for safety, precautions for the information you input.

*The techniques here are a summary of general methods widely recommended in the public guides and research of various companies (OpenAI, Anthropic, Google, etc.). Their effect varies by model, task, and input, and results are not guaranteed.

1. What is prompt engineering?

Prompt engineering is the skill of designing and improving your instruction (the prompt) so that an AI — ChatGPT, Claude, Gemini, and so on — returns the answer you want. It is not difficult programming; it is closer to "the craft of how you say things."

Why does it work? Today's AI (a large language model, or LLM) builds an answer by predicting, "plausibly," how the given text should continue. In other words, change the words at the entrance (the prompt), and the answer at the exit changes greatly. For the very same question, simply adding a role, context, or output format can sharply raise accuracy and usability — that is the power of prompt engineering. No special talent is needed. Learn the "patterns," and fix things a few times. That alone makes anyone better.

2. The three principles that change your results

Before the fine techniques, lock down just three principles that everything rests on. Miss these, and any technique you add will spin its wheels.

One more humble but effective trick: write "do X" rather than "don't do Y." With prohibitions alone, the AI cannot tell what to do instead. Rather than "don't use jargon," say "use words a middle-schooler would understand" — pointing to the direction you want is more stable. It is an iron rule emphasized again and again in the major guides.

3. [Core] The 6 parts of a good prompt

Let us turn the 6 parts from the opening diagram into actual writing. You do not need all six every time. Choosing the parts a task needs, and filling in what is missing, is how pros use it.

Combine them and you get a prompt like this. The more parts you supply, the more steadily the reply lands where you aimed.

Compare that with "tell me a dinner I can make in 15 minutes," and you can imagine how much more usable the answer will be. The more parts you add, the easier it is for the AI to fit "your situation."

4. 7 practical techniques that work right away

With the 6 parts in mind, here are 7 concrete techniques you can use tomorrow. Each helps even on its own.

Of these, the highest return comes from ⑦, iteration. The essence of prompt engineering is not "writing one perfect sentence," but steering the answer closer through dialogue with the AI. Treat the first reply as a rough draft, and the whole thing gets easier — and you get better.

5. Before → After, with a real example

Let us put the principles and techniques side by side in a common scenario — a "bad example" versus a "good example." The difference is obvious at a glance.

The point is that the After uses no special expertise. It just adds "who, to whom, what, in what form, in what tone." That difference is what shapes your daily working time and the quality of the output.

6. Next-level techniques (CoT, chaining, more)

Once you are comfortable with the basics, knowing some "advanced moves" that shine on complex tasks gives you an edge. Just grasp the idea — that is enough.

One important update to add. "Reasoning models" — such as OpenAI's o-series and Claude's extended thinking — are built to perform CoT internally and automatically. So with these models, the need to write "reason step by step" every time has dropped. If anything, it works better to clearly convey "what you want to achieve (the goal)." Techniques are not universal; the best move changes with the model you use — keep that sense, and you will level up.

7. 7 common mistakes

Let us head off the familiar pitfalls that block your progress. If any ring a bell, that is your room to grow.

• Dumping it vaguely: "Make it nice" gives the AI no standard. Be specific about conditions.
• Zero context: Skipping background, goal, and audience. Remember the AI does not know your situation.
• Prohibitions only: Just "don't do Y." Instead, point the way with "do X."
• Cramming too much at once: Five requests in one prompt. Splitting improves both accuracy and fixability.
• No format specified: Leaving the output shape to the AI. Decide table, length, and tone first.
• No examples: A "pattern" words cannot fully capture lands instantly if you show one model.
• Giving up after one shot: Judging on the first answer. If you assume you will add and fix, you can push it further.

Flip these around, and you improve fast. "Vague → specific," "omitted → context," "dumping → splitting," "one shot → iterate." Those four arrows are all you need to keep in mind.

8. Model-specific tips and safety

Finally, two things that often trip people up in practice — the differences between models, and safety when you input.

Model tendencies: ChatGPT, Claude, and Gemini each have slightly different strengths and quirks. People say Claude is good at organizing long text and polished writing, Gemini at fresh information and search integration, and ChatGPT at all-round balance (using each by task is wise). For which to pick, see ChatGPT vs Claude vs Gemini comparison. But across all of them, the 6 parts in this article work.

Safety: As important as getting better at prompts is "what you are allowed to input." Carelessly pasting confidential information, personal data, or internal materials risks data leakage and policy violations. If you use AI at work, be sure to read precautions for the prompts and information you give AI. Good prompting and safe input are two wheels of the same cart.

Summary

Here are the practical points of prompt engineering, condensed.

• The essence: The skill of designing and improving your instruction to AI. Not code but "the craft of how you say things" — anyone can improve starting today.
• Three principles: ① be specific ② give context ③ specify the output. That is the whole foundation.
• The 6 parts: Role, context, instruction, examples, format, constraints. Not all every time — fill in what is missing.
• 7 techniques: Role / model / step-by-step / fixed format / structure / split / iterate. The strongest is "iterate."
• Advanced: CoT, self-consistency, chaining, ReAct. With reasoning models, stating the goal works well.
• Two wheels: Good prompting + safe input. Beware of pasting confidential data.

In the end, prompt engineering is not "a special person's skill." With just the attitude of "turn vague into specific, turn dumping into dialogue," your AI will look transformed from today. Start by adding one "role" and one "output format" to your usual instruction in the ChatGPT in front of you. For concrete request examples, prompt tips for asking AI to build an app is also practical.

FAQ