Good Prompts vs. Bad Prompts: Comparing the Same Request

Frequently asked questions

A closer look

Why do bad prompts only get obvious answers, and what do good prompts contain?

Because they leave too many choices open. A sentence like "make me some materials" or "give me something tasty" says nothing about who will read it or how long it should be. It's like saying "go anywhere" without any route coordinates. So the AI fills in the blanks in the safest way. That's why the result turns out obvious.

There's one more thing. Conversational AI creates text by picking a plausible next word rather than searching for the "truth." So even when a question lacks supporting evidence, it can sometimes state untrue things with confidence. This is called hallucination (when AI makes up plausible-sounding falsehoods). A bad prompt leaves this risk wide open too.

A good prompt, on the other hand, is closer to a work request than a single question. When the reader, purpose, content scope, tone, length, and output format are included, the range the AI has to work through gets smaller. An easy way to picture it: telling a colleague "make me some materials" is worse than saying "summarize market size, competitors, and risks in a 5-slide deck for tomorrow's executive briefing."

What changes when you write the same request two ways?

Let's compare using a blog post request. The bad version is a single line: "Write a blog post about using AI." The good version looks like this.

"Write a blog post for office workers who are new to generative AI. The goal is to make them want to try using AI at work. Keep technical explanations to a minimum and include examples such as drafting emails, summarizing meeting notes, and researching information. Make it around 1,200 characters, in a friendly tone that doesn't exaggerate. At the end, include 3 example prompts they can try right away."

Salesforce's comparison points the same way. "Write a promotional email for our company's new CRM" tends to end up generic. Adding the writer's role (copywriter), the recipient (IT team lead), the format (bullet points), and a constraint (no technical jargon) gets you closer to a draft you can send without edits, according to the explanation. You can read the original at the Salesforce prompt guide.

What does a copy-and-paste template with a role look like?

A commonly used framework is role, purpose, format, and example. Here, the role isn't a grand title like "you are the world's best expert." It's there to tell the AI which perspective you want. The Korea Tech University AI guide page introduces C.R.E.A.D as a basic formula, and adds that the example sentences are for reference, so you should keep testing and refining them.

Just swap out the blanks below. "Write a [type of output] for [reader]. The goal is [what you want to get from this piece]. It should cover [scope], with a [mood] tone and a length of [amount]. Use a [table/bullets/paragraphs] format. If something isn't in the material I gave you, don't make it up; answer 'cannot verify'."

The key is the last sentence. Telling the AI not to guess at things it doesn't know and to use only the material you provided helps reduce hallucination. But writing it once doesn't remove the limitations, so double-check numbers and dates yourself.

If I don't like the first answer, should I rewrite everything from scratch?

No. Since a conversation carries context forward, it's faster to give follow-up instructions like "Make this part more specific" or "Make the tone more formal." Salesforce calls this Iterative Refinement, and encourages you to let go of the pressure to write it perfectly in one line. Instead of trying to set the route perfectly at once, you make small course corrections.

There's a knack to length as well. The Moket Prompt Bible explains that Korean tends to be split into more tokens (the units AI breaks text into when reading) than English, so it helps to drop greetings and humble phrasing and keep the key points concise. There's also advice that AI tends to miss the middle of long text, so you should put the instructions you must have followed at the very beginning or the very end of the prompt. These two points come from that source's explanations, so try them yourself to check. Moket AI Prompt Bible, Chapter 1

Are there things that still won't work even with a good prompt?

Yes. A good prompt only improves the result; it can't turn wrong information into fact or fill in data that doesn't exist. It's like ordering food: you can get "a menu that's not spicy and has no meat," but you can't make the restaurant cook with ingredients it doesn't have.

At work, you also need to be careful about what you enter. The basic rule is not to put company secrets or personal information in as they are, and prompt management and security are becoming issues for companies. This is also covered in Tilnote's prompt explanation.

Also, image, video, and music AI have their own grammar. Video models like Kling 3.0 respond to scene descriptions that specify movement and camera direction rather than photo captions (Morphic Kling 3.0 guide). For Suno, the guide explains that it works better with tags such as genre, vocals, and tempo, along with structure tags like [Verse] and [Chorus] (Moket Suno Prompt Master). Rather than carrying over what you learned from conversational AI as is, check each tool's guide first.