How to Write Better AI Prompts: A Practical Guide for ChatGPT, Claude and Gemini

Written by

in

You do not need to be technical to write great prompts. You need to give an AI the same information you would give a smart new colleague on their first day: what the job is, who it is for, what good looks like, and what to avoid. This guide explains the techniques that consistently improve results across ChatGPT, Claude and Gemini, with examples you can copy and adapt.

The anatomy of a great prompt

Nearly every high-quality prompt has the same six ingredients. You will not always need all six, but checking for them is the fastest way to fix a disappointing answer.

  1. Role: who the AI should act as. “A CPA who works with freelancers” gives different answers than “an assistant”.
  2. Task: the specific thing to produce, stated with a strong verb: write, compare, summarize, critique, plan.
  3. Context: facts the AI cannot guess, such as your product, your audience, your past attempts and your constraints.
  4. Requirements: tone, length, format and rules such as “no jargon” or “under 150 words”.
  5. Examples: a sample of what you want, or of your own style. This is often the single most powerful addition.
  6. Output format: exactly how to present the answer, such as a table, bullet list, JSON or email with subject line.

The Prompt Enhancer walks you through these ingredients and assembles the prompt for you.

Before and after

Weak: Write a blog post about remote work.
Strong:
You are an HR consultant who advises US companies with 20-200 employees.

Write a 900-word blog post for small-business owners who are deciding whether to bring remote employees back to the office.

Context: our readers are practical and skeptical of trends. Many have tried hybrid schedules with mixed results.

Requirements:
- Balanced: cover real advantages and drawbacks of each option
- Include one short, realistic example of a 50-person company
- Plain English, no buzzwords
- H2 subheadings, and end with a 5-question checklist owners can use to decide

Do not cite specific statistics unless you are confident they are accurate; if you mention research, describe it generally.

The second prompt is longer, but every line adds information or removes a likely problem. That is the test of a good prompt: not length, but density of useful instruction.

Seven techniques that work in every major model

1. Show an example (few-shot prompting)

If you want a specific style or format, show it. Two or three examples teach the AI more than a paragraph of description.

Write 5 product taglines in this style:
- "Coffee that keeps up with you."
- "Shoes for the long way home."
Product: a reusable water bottle that keeps drinks cold for 24 hours.

2. Ask for reasoning before the answer

For analysis, math, planning or decisions, ask the model to work through the problem before concluding. Many current models do this automatically, but asking explicitly still helps on complex tasks, and lets you check the logic.

Before recommending an option, list the key factors, evaluate each option against them, and then give your recommendation with the single biggest risk.

3. Let the AI ask questions first

Before you start, ask me up to 5 questions about anything you need to know to do this well. Wait for my answers.

This is ideal for complex tasks like business plans, where you might not realize what information is missing.

4. Use delimiters for pasted material

When you paste documents, separate them clearly from your instructions. Claude responds especially well to XML-style tags; ChatGPT and Gemini work well with tags or Markdown headings.

Summarize the contract below in plain English and list any unusual clauses.

<contract>
[Paste Contract]
</contract>

5. Say what to do, not only what to avoid

“Do not be wordy” is weaker than “Use short sentences and keep it under 120 words.” Positive instructions give the model a target.

6. Break big tasks into steps

Instead of asking for a complete marketing plan in one prompt, ask for a customer analysis first, then positioning, then the channel plan. Review each step before moving on. This is called prompt chaining, and it produces much better long work.

7. Ask for a self-review

After writing your draft, review it against my requirements, list anything missing or weak, and then give me only the improved final version.

Model-specific tips

  • ChatGPT (OpenAI): responds well to Markdown structure and clear headings. Use Custom Instructions or Projects to save context you repeat often.
  • Claude (Anthropic): excellent with long documents and detailed instructions. XML tags help it keep sections separate. Projects let you store reference files.
  • Gemini (Google): integrates with Google search and Workspace. Ask it to flag information that may be out of date, and verify facts it pulls from the web.
  • Image models such as Midjourney: use descriptive phrases, not instructions: subject, style, lighting, composition, then parameters like –ar. See our product photography prompts.

Common mistakes

  • Too little context. The AI fills gaps with generic assumptions. Add specifics.
  • Several tasks in one prompt. Split them.
  • Accepting the first answer. Reply with specific feedback: “Make the second paragraph more concrete and cut the last one.”
  • Trusting facts blindly. AI can state wrong information confidently. Verify numbers, quotes, laws and citations.
  • Pasting sensitive data. Remove personal, financial and confidential information first.

Turn good prompts into templates

When a prompt works, save it with the changing parts in square brackets, like [Audience] and [Topic]. Paste it into our Prompt Template Builder to get a reusable fill-in form.

Frequently asked questions

Is prompt engineering still useful as models get smarter?

Yes. Smarter models need fewer tricks, but they still cannot read your mind. Context, examples and clear requirements will always matter because they carry information only you have.

How long should a prompt be?

As long as it needs to be to include the necessary information, and no longer. A quick question needs one sentence. A brand-specific marketing email might need 200 words of context.

Do magic phrases like “take a deep breath” work?

Effects from phrases like this were small and inconsistent even in older models. Clear context and requirements make a far bigger difference.