Two people can use the same AI assistant and get completely different results. The difference is rarely the subscription tier — it is how the request is written. These five principles work across every major assistant and take about ten minutes to learn.
Why vague prompts produce vague answers
A model answers within the context it is given. With little context, the safest output is a generic answer that fits almost any situation, which is exactly what feels useless. Supply constraints and the answer narrows to your case.
Good prompting is not about magic phrasing. It is about removing the room the model has to guess wrong.
1. Give a role and a situation
The assistant does not know who you are or why you are asking. Supplying role, audience and purpose changes the output more than any other single fix.
| Weak | Strong |
|---|---|
| Write a product description | You are a copywriter. Write a 300-character description of a sugar-free sparkling water for health-conscious professionals in their 30s. Confident tone, no exaggerated claims. |
| Plan a trip | Plan a 3-day Kyoto itinerary for two people in their 60s. Maximum one walking-heavy site per day, vegetarian-friendly meals. |
| Fix this code | I am new to Python. Explain why this code raises an error in plain language, then show the corrected version. |
2. Specify the shape of the output
If you do not state the format, the model picks one. Three parameters cover most cases:
- Length: under 200 words, three paragraphs, exactly 10 bullet points
- Format: as a comparison table, as an email, as a numbered checklist
- Tone: explain it to a beginner, formal enough for a board update, conversational
Asking for a table is unusually effective. Building one forces the model to structure its thinking, which makes vague filler harder to produce.
3. Show an example
One example beats a paragraph of description. Paste a headline you liked, a previous report section, or a writing sample and say to match that style. For recurring work with a fixed format — meeting notes, weekly updates, product listings — save the example and reuse it.
4. Iterate instead of perfecting the first prompt
Treat the first response as a draft. Refining through conversation is faster than engineering a perfect opening prompt:
- Make the second section more concrete.
- Give me five more headline options.
- Point out the weakest claim in what you just wrote.
- Argue the opposite position.
One exception: if the conversation has gone badly off course, start a new one. Bad context tends to persist.
5. Build verification into the request
Because models produce confident errors, put the checking instruction in the prompt itself:
- Flag anything you are not confident about.
- Include source links — then actually open them, because fabricated links are common.
- List separately every claim in this answer that should be fact-checked.
None of this eliminates errors. Dates, figures, names, and legal or medical statements still need verification against a primary source.
A template you can copy
Role: You are a [role].
Situation: [context, audience, constraints]
Task: [what you want]
Format: [length, structure, tone]
Rules: Flag uncertainty. Exclude [what you do not want].
Frequently asked
Do I get better answers in English?
For everyday tasks in widely spoken languages the gap has narrowed substantially. For narrow technical fields, English training data is still deeper, so re-asking in English is worth trying when an answer feels thin.
Should prompts be long?
Length is not the goal. The test is whether someone reading only your prompt could do the task. Extra words that add no constraint make the request worse, not better.
Does telling it to think step by step still help?
Less than it used to. Current reasoning models already do this internally. A more useful instruction now is show the reasoning behind your conclusion, which lets you check the logic rather than just the answer.
Prompting is a habit more than a skill. Apply these five deliberately for a week and they stop requiring conscious effort.
작성자 소개
