Ask This Way to Get Better Answers From ChatGPT, Claude and Gemini

What You Get

Type a plain "Help me write..." or "Explain..." to ChatGPT, Claude, or Gemini and you'll get an answer, but it's often only an average one. Spend another ten seconds and ask according to the five rules below, and the same model can give you much better results. This article explains each of the five rules, and each comes with wording you can copy directly.

This image is a schematic diagram we drew.

Rule 1: Set the Role First, Then Add Constraints

In the opening sentence, first tell it "who you are." When the identity changes, its approach and word choice change with it. Then add specific constraints:

You are a consultant who is skilled at brand marketing. Please answer in this role.
Requirements: under 800 words; no jargon; state the conclusion first, then the reasoning, and finish with one example.

"Just write something" is the worst instruction. The clearer the framework, the more stable the answer.

Rule 2: Describe the Goal as the "Finished Result"

Don't only tell it "what to do"; spell out what the reader will experience once it's done.

• Weak: Help me write a Weibo post.
• Better: Write a Weibo post that makes people want to act immediately after reading it.

There's also a trick: show it counter-examples.

Don't use this kind of phrasing: "Shocking! You'll never guess..."
Don't write in this tone: exaggerated, with exclamation marks piled up.

Give two or three examples it must not imitate, and the off-target answers will drop considerably.

Rule 3: Don't Hand It All the Work at Once

Writing usually involves these steps: pick a topic, outline, write the body, revise the draft. If you have the AI do everything in one go, quality usually drops. Split it up and ask step by step:

1. First have it give you 5 topic options, and you pick one.
2. Then have it outline that topic, and you revise the outline.
3. Then have it write the body according to the outline.
4. Finally have it revise the draft according to your comments.

Review the result of each step before starting the next.

Rule 4: Add a Line Against Made-Up Facts

Models sometimes invent numbers, dates, person names, and organization names with a straight face. Add this to your prompt:

For any information that has not been verified, mark it "unverified." Do not mix guesses with facts.

This line noticeably reduces fabrication, but it cannot guarantee that there are no errors. For numbers, dates, person names, and organization names, especially important ones, you must check them yourself.

Rule 5: Don't Stop at the First Version

Have it score its own answer, then revise:

Score your previous answer (out of 10), name 3 things that could be improved, then rewrite a version based on those improvements.

Most people stop after one round of asking. Following up two or three more rounds is often what gets you a version you can actually use.

One Combined Template

Notes

• These five habits are general prompting practices that work in ChatGPT, Claude, and Gemini, though results will vary between models.
• You don't need to use all five every time. For simple questions, one or two sentences are enough; the full set is worth using for complex work like writing articles, drafting plans, or analyzing materials.
• For content involving medical, legal, or financial matters, treat AI answers as reference only and confirm with a professional.