ChatGPT vs Claude vs Gemini: how to adapt your prompts to each

A good prompt works in ChatGPT, Claude and Gemini with very few changes, because all three follow the same basics: a clear role, a specific task, real context and an explicit output format. The differences show up at the edges: how each reads structure, how it handles a 60-page document, how it treats formatting, and where you store instructions so you stop retyping them. This guide covers the shared core first, then the practical adjustments for each assistant, then one portable template you can paste into any of them.

What stays the same in all three

Before tuning for a specific model, get the fundamentals right. Most bad results in any assistant come from missing information, not from picking the wrong one.

Every strong prompt, in every assistant, contains:

• Role: who the model should act as ("a B2B copywriter who writes for finance teams").
• Task: one clear job with a verb ("write", "compare", "rewrite", "extract").
• Context: audience, goal, background facts, what you already tried.
• Format: length, structure, headings or not, bullet points or prose.
• Constraints: what to avoid, tone, things the model must not invent.

All three vendors say versions of the same thing in their own documentation: be direct, be specific, separate instructions from the material they apply to, and show an example when the format matters. If you are new to this structure, start with prompt engineering for beginners, then come back for the model-specific tweaks.

Here is a prompt that fails everywhere, and a version that works everywhere.

Before:

Summarize this report.

After:

You are an analyst preparing a briefing for a busy operations director.
Summarize the report below in 5 bullet points, each under 25 words.
Focus on decisions she needs to make and any numbers that changed since last quarter.
If the report does not state something, say "not stated" instead of guessing.

<report>
[paste report]
</report>

The after version works in all three assistants because it removes the guesswork: who is reading, how long, what matters, and what to do with missing information. The tags around the report tell the model where your instructions end and the material begins.

Structure: XML tags, Markdown and delimiters

All three models benefit from clearly separated sections. The style of separator is where small preferences appear.

• Claude: Anthropic's prompting best practices recommend XML tags such as <instructions>, <context> and <input> to separate the parts of a complex prompt. Claude handles them especially well, and you can ask it to put its answer inside a tag too.
• Gemini: Google's Gemini prompting guidance recommends clear delimiters, either XML-style tags or Markdown headings, and suggests picking one style and using it consistently within a prompt.
• ChatGPT: OpenAI's guidance has long suggested delimiters such as triple quotes, XML tags or section titles to mark which text is instructions and which is data. Markdown headings work well here.

Practical rule: XML-style tags are the safest portable choice. All three understand them, and they never get confused with the content you paste, which may already contain Markdown.

Length and level of detail

None of the three can read your mind about length, and default length varies between models and even between versions of the same model. Treat length as something you always specify.

• Give a number: "under 150 words", "5 bullets", "3 short paragraphs".
• Say who it is for: "for someone who will read this on a phone between meetings".
• If you want more than the minimum, ask for it. Anthropic's docs note that current Claude models follow instructions precisely, so if you want "above and beyond" output, say so instead of expecting the model to infer it.

Before:

Explain how our refund policy works for the support team.

After:

Explain our refund policy for new support agents in their first week.
Format: a 6-step numbered checklist an agent can follow during a live chat, then 3 example customer questions with the correct answer.
Keep each step under 20 words. Use the policy text below and nothing else.

<policy>
[paste policy]
</policy>

The after prompt sets a format and a length that are hard to misinterpret, so ChatGPT, Claude and Gemini produce output of roughly the same shape. That consistency matters if you switch between assistants or compare their answers.

Formatting: getting the output you actually want

Each assistant has its own default habits around headings, bullets and bold text. The fix is the same in all three: state the format explicitly and describe what you want, not only what you do not want.

• Say "write in flowing paragraphs with no headings" rather than only "don't use bullet points".
• Anthropic notes that the style of your prompt can influence the style of Claude's reply: a prompt full of Markdown tends to produce Markdown, and a plain prose prompt tends to produce prose. This is worth trying in any assistant.
• If the output goes into another tool (an email client, a CMS, a spreadsheet), say so: "plain text, no Markdown symbols, because I will paste this into Gmail".
• When format matters a lot, give one short example of the output. A single example is often more reliable than a paragraph of description. See few-shot prompting for how many examples to use and how to format them.

Long documents and big context

All three assistants can now work with long inputs, and many Gemini models advertise especially large context windows, on the order of a million tokens or more according to Google's docs. A bigger window does not remove the need for good structure, though. Two habits help everywhere:

1. Put the document first and the question last. Anthropic recommends placing long documents near the top of the prompt, above the instructions and query. Google gives similar advice for Gemini: put your specific question at the end, after the material.
2. Restate the key instruction after long material. For ChatGPT especially, repeating your main instruction after a very long paste is a simple safeguard against it drifting toward a generic summary.

A layout that works in all three:

<documents>
<document source="Q3 board deck">[paste]</document>
<document source="Customer survey notes">[paste]</document>
</documents>

Using only the documents above, answer: [your question].
Quote the exact sentence you relied on for each claim, with its source name.
If the documents do not answer it, say so.

Asking for quotes is useful in every assistant because it makes the model ground its answer in your material and lets you check it quickly.

Where to save instructions in each assistant

Retyping the same role and rules into every chat is the biggest time sink. Each assistant has a place to store them.

ChatGPT

• Custom instructions: set in Settings under Personalization. They apply across your chats, so keep them general: who you are, how you like answers formatted, what to avoid.
• Projects: group chats and files around one piece of work, with project instructions that apply only inside that project. According to OpenAI's help center, project instructions override your general custom instructions inside that project.
• Custom GPTs: a reusable assistant with its own instructions and knowledge files, built in the web app, on plans and workspaces that allow creating them.

Claude

• Profile preferences: account-wide instructions in your settings that apply to all your conversations.
• Projects: each project has a knowledge base for files and its own instructions, applied to every chat inside the project. Note that chats in the same project do not share context with each other unless the information is in the project knowledge.
• Styles and skills: Claude has offered ways to save tone and format preferences, such as writing styles built from your own samples. The exact options shown depend on your plan and app version, so check what your Claude account offers.

Gemini

• Gems: a saved custom version of Gemini with its own instructions and optional knowledge files. Google's Gems help page suggests covering persona, task, context and format in the instructions, and offers a button to have Gemini rewrite a rough description into fuller instructions. Remember to click Save: previewing a Gem does not save it.

The same text can go into all of these. A well-written block of standing instructions is, in practice, a system prompt. For complete examples you can paste into any of these fields, see these system prompt examples.

Small tuning tips per assistant

Once the core prompt is solid, these adjustments are worth trying:

• ChatGPT: use Markdown headings or triple-quoted blocks to split long prompts, restate the main instruction after big pastes, and use a Project when you return to the same work over several days.
• Claude: wrap each part in XML tags, put documents above instructions, explain why a rule exists ("keep it under 100 words because it is read on mobile"), and use calm wording. Anthropic notes that some recent Claude models are more responsive to system prompts than earlier ones, so all-caps warnings like "CRITICAL: YOU MUST" can cause overcorrection.
• Gemini: be direct and concise, include a short example of the output format, keep your delimiter style consistent, and place your question at the very end of long prompts. Use a Gem for any task you repeat weekly.

One portable template for all three

This template follows the shared advice from all three vendors: role up front, material separated by tags, explicit format, question at the end.

<role>
You are [role with relevant expertise], helping [who I am] with [type of work].
</role>

<context>
[Background: audience, goal, what has been tried, why this matters.]
</context>

<material>
[Paste documents, data or drafts here. Leave empty if none.]
</material>

<task>
[One clear instruction with a verb, e.g. "Rewrite the landing page intro above for first-time visitors".]
</task>

<format>
- Length: [e.g. under 200 words / 5 bullets]
- Structure: [e.g. headline, then 3 short paragraphs]
- Style: [tone in 2 to 3 adjectives], avoid [phrases, jargon, emoji]
</format>

<rules>
- Use only the material above for facts. If something is missing, say "not provided".
- Ask me up to 3 questions first if the task is ambiguous.
</rules>

If you want this structure without filling it in by hand, a prompt generator such as Promptay turns one sentence into this role, task, context, format and constraints layout, tunes it for ChatGPT, Gemini or Mistral, and lets you open it directly in ChatGPT, Claude or Gemini.

Frequently asked questions

Do I need different prompts for ChatGPT, Claude and Gemini?

Mostly no. A prompt with a clear role, task, context, format and constraints works well in all three. The differences are small adjustments: XML tags suit Claude especially well, Gemini responds well to direct wording and a consistent delimiter style, and ChatGPT benefits from restating key instructions after long material. Fix the fundamentals first, then tune.

Which AI is best for long documents?

All three handle long documents, and many Gemini models advertise especially large context windows. In practice, structure matters as much as window size: put the document first, the question last, and ask the model to quote the passages it used. That makes any of the three more accurate and easier to check.

Do XML tags work in ChatGPT and Gemini, or only in Claude?

They work in all three. Anthropic recommends them specifically for Claude, Google lists XML-style tags as one of the delimiter options for Gemini, and OpenAI has long suggested XML tags or similar delimiters to separate instructions from data. Tags are the most portable way to structure a complex prompt.

What is the equivalent of a custom GPT in Claude and Gemini?

In Gemini, the closest equivalent is a Gem: a saved assistant with its own instructions and knowledge files. In Claude, the closest equivalent is a Project, which combines a knowledge base with project instructions that apply to every chat inside it. All three store what is effectively a system prompt plus reference files.

Why does the same prompt give different answers in each assistant?

Each model was trained differently and has its own defaults for length, formatting and tone. The more decisions your prompt leaves open, the more those defaults show. Specifying length, structure, audience and what to do with missing information narrows the gap and makes answers far more consistent across assistants.