How to Write Better Prompts for ChatGPT, Claude and Gemini

Models keep improving. In the last few months alone we got GPT-6 in ChatGPT, Claude Sonnet 5 and Opus 5.5, and a steady run of new Gemini releases. Yet the most common complaint we hear hasn’t changed: “The answer was generic.” Usually the model isn’t the problem. The prompt is.

The good news is that OpenAI, Anthropic and Google all publish prompting guides, and their advice overlaps a lot. This guide boils that shared advice down into six steps that work in all three chatbots, with before-and-after examples for each.

Key takeaways

• Write prompts the way you would brief a smart new colleague. Explain the goal, who it’s for, any constraints, and what a good result looks like.
• Use four building blocks: role, task, context and format. Google’s Workspace guide uses almost the same framework.
• Show, don’t just tell. One or two examples of the output you want beat a paragraph of adjectives.
• Keep instructions separate from your material. Delimiters or XML-style tags stop the model from mixing them up.
• Iterate. Treat the first reply as a draft, then tighten your instructions.

Why prompts matter more than you think

A language model can only work with what you give it. It can’t see your audience, your deadline, your brand voice or the five drafts you already rejected. When information is missing, it fills the gaps with the safest, most average guess. That’s where the generic tone comes from.

Anthropic’s documentation sums this up neatly: “If a colleague would be confused by your prompt, Claude will be too.” That test works just as well for ChatGPT and Gemini.

Step 1: Give the model a role

Opening with one line that sets a role focuses the vocabulary, depth and tone. Anthropic recommends it, and Google’s Workspace guide lists “persona” as one of its four parts of a good prompt.

Before:

Explain our new expense policy.

After:

You are an HR communications specialist who writes in plain, friendly English.
Explain our new expense policy to employees who have never filed an expense report.

The role on its own won’t get you far. It works because it tells the model who is speaking and who is listening.

Step 2: Be specific about the task and the outcome

Vague words like “short,” “better” or “professional” mean something different to everyone. OpenAI’s guide suggests replacing them with concrete instructions, such as a sentence count instead of “fairly short.” Anthropic adds that if you want the model to go beyond the basics, you have to ask for that explicitly.

Before:

Write a short LinkedIn post about our product launch.

After:

Write a LinkedIn post announcing the launch of our invoicing app for freelancers.
- Length: 80-120 words
- Audience: freelance designers and developers
- Goal: get readers to sign up for the free trial
- Include one concrete benefit (invoices sent in under 60 seconds)
- End with a single clear call to action
- No hashtags, no emojis

Step 3: Add context, and explain why

Context is the background the model can’t guess: what happened before, what you’ve already tried, and which constraints matter. Anthropic also recommends explaining the reason behind an instruction, because that helps the model apply it sensibly in situations you didn’t anticipate.

Before:

Don't use any jargon.

After:

This summary goes to our board, and half the members have no technical background.
Avoid technical jargon; if you must use a term like "API", explain it in one short clause.

The second version tells the model what to avoid and also why. So when it hits a borderline term you never mentioned, it can make a sensible call.

Step 4: Separate your instructions from your material

When you paste in an email thread, report or transcript, mark clearly where it starts and ends. OpenAI suggests putting the instructions first and using delimiters such as triple quotes or ###. Anthropic recommends XML-style tags. For long documents it suggests placing the material near the top and your question at the end, which it says can noticeably improve answer quality on complex inputs.

Before:

Summarize this and tell me what to reply [pastes 40-message email thread]

After:

<thread>
[paste the email thread here]
</thread>

You are my executive assistant. Using only the thread above:
1. Summarize the open decisions in 3 bullet points.
2. List anyone waiting on a reply from me.
3. Draft a reply to Sara under 100 words that confirms the Friday deadline.

These tags work fine in ChatGPT and Gemini too. The models don’t need special syntax. They just need clear boundaries.

Step 5: Show an example of what good looks like

Examples are one of the most reliable ways to control format and tone. Anthropic suggests three to five varied examples for important tasks. In everyday chat, even one or two make a big difference. OpenAI recommends starting without examples and adding them only if the first results miss the mark.

Before:

Write product descriptions for these three mugs.

After:

Write product descriptions for the three mugs listed below.
Match the style of this example exactly:

<example>
The Dawn Mug - Hand-glazed stoneware that holds 350 ml of your
favourite brew. Dishwasher safe, heavy enough to feel good in
the hand, and cheerful enough to start any Monday.
</example>

Mugs:
1. Forest Mug, green glaze, 400 ml
2. Slate Mug, matte grey, 300 ml
3. Tide Mug, blue speckle, 350 ml

Step 6: Say what you want, not just what you don’t

Both OpenAI and Anthropic make this point. A list of bans tells the model what to avoid but not what to aim for, so it’s left guessing. Turn each “don’t” into a “do.”

Before:

Don't use bullet points. Don't be too formal. Don't make it long.

After:

Write two short paragraphs of flowing, conversational prose,
as if you were explaining this to a friend over coffee.

Putting it all together: a reusable template

Here is a template that combines all six steps. Save it somewhere handy and fill in the brackets.

ROLE: You are a [role] who [relevant expertise or style].

TASK: [Exactly what you want produced, in one or two sentences.]

CONTEXT: [Who it's for, why it matters, what's already been tried,
any constraints - and the reason behind each constraint.]

FORMAT: [Length, structure, tone, headings or bullets, language.]

EXAMPLE (optional):
<example>[A sample of the output you want]</example>

MATERIAL (optional):
<material>[Pasted text, data or notes]</material>

Before you finish, check your answer against the format rules above.

The last line comes from Anthropic’s guide. Asking the model to check its own work against your criteria catches surprisingly many errors, especially in calculations and code.

Tips for each chatbot

ChatGPT

• Use Projects or custom instructions to save context you repeat often, such as your role, company or writing style, so you don’t paste it into every chat.
• For research-heavy questions, turn on Deep Research and give it a scoped brief, not a one-line question.
• With GPT-6’s Intelligent UI, ChatGPT may answer with charts or small tools. If you want plain text only, say so in your prompt or adjust the setting.

Claude

• Claude handles long documents and XML-style tags especially well. Wrap each pasted document in its own labelled tag.
• Upload reference files to a Project so every new chat starts with the same background.
• Anthropic warns that recent models react strongly to forceful wording. ALL-CAPS commands like “CRITICAL: you MUST” can make the model over-apply a rule, so calm, clear instructions work better.

Gemini

• Google’s Workspace guide uses the same four-part structure: persona, task, context, format. In its testing, the most successful prompts averaged around 21 words, so a couple of well-chosen sentences can be enough.
• In Gmail, Docs and Drive you can point Gemini at existing files instead of pasting them, which is faster and keeps formatting intact.

Iterate like an editor

Even a great prompt rarely gets you a perfect answer on the first try. Treat the reply as a draft and give feedback the way you would to a colleague:

1. Be specific about what’s wrong. “Paragraph two is too salesy, rewrite it as neutral advice” works much better than “make it better.”
2. Ask for options. “Give me three alternative headlines with different angles” helps you see what’s possible.
3. Ask for action, not suggestions. Anthropic notes that “Can you suggest changes?” tends to get you suggestions, while “Rewrite this paragraph to…” gets you the actual edit.
4. Start fresh when a chat drifts. If a long conversation goes off track, copy your best prompt into a new chat and add what you’ve learned.

Common mistakes to avoid

• Asking several unrelated questions at once. Split big jobs into steps and check each one.
• Trusting facts without checking. Better prompts reduce errors but don’t remove them. Verify numbers, quotes and citations.
• Pasting sensitive data into personal accounts. Use your organization’s approved workspace for anything confidential.
• Overloading the prompt. More context helps, up to a point. Cut instructions that don’t change the result.

The bottom line

Good prompting isn’t a secret trick. It’s clear communication. Give the model a role, a specific task, the context and the reasons behind it, a clear format, and an example when the format matters. Do that consistently and you’ll get better results from ChatGPT, Claude and Gemini than most people ever see.