ChatGPT vs Claude vs Gemini: how prompts differ

ChatGPT, Claude and Gemini all understand plain, well-organized instructions, so a good prompt will work in any of them. But each provider publishes its own prompting advice, and the small differences (how to mark sections, where to put long documents, how system instructions work) add up when you use them every day. This guide compares the three, based on the providers' own documentation, then shows the same task prompted three ways. Models update often, so treat this as a guide to prompting style rather than a benchmark of which assistant is "smarter".

What works the same everywhere

Before the differences, the common ground. All three providers recommend the same fundamentals:

• Be explicit. State the task, the audience and the success criteria. None of them read your mind.
• Give context. Explain why you need something and who it is for; the answer adapts to that purpose.
• Separate instructions from material. Use headings, tags or delimiters so pasted content is not confused with your request.
• Show examples when a format or style matters.
• Iterate with specific feedback instead of starting over.

If your prompt already does these things, the model-specific adjustments below are refinements, not requirements. Our guide on how to write AI prompts covers the fundamentals in depth.

ChatGPT: Markdown structure and explicit instructions

OpenAI's prompting guides recommend using Markdown and XML tags to show the structure of a prompt: Markdown headings for sections and hierarchy, XML-style tags or triple quotes to wrap documents and user-provided content. Recent OpenAI models are described as following instructions precisely, which cuts both ways: clear instructions are followed closely, and contradictory ones cause confusion. Review long prompts for conflicts.

Practical habits that suit ChatGPT:

• Organize longer prompts with ## Role, ## Task, ## Output format style headings.
• Put stable preferences (your role, preferred tone, language) into custom instructions or a Project, instead of repeating them in every chat.
• For reasoning-focused models, state the goal and constraints and skip "think step by step"; OpenAI's guidance says those models reason internally.

Claude: XML tags and documents first

Anthropic's documentation explicitly recommends XML tags to structure prompts for Claude: wrap instructions, context, examples and documents in descriptive tags such as <instructions>, <context> or <document>. There is no special list of tag names; consistency is what matters. You can also ask Claude to put parts of its answer inside tags, which makes it easy to extract them.

For long inputs, Anthropic advises placing the long documents near the top of the prompt, above your instructions and question, and wrapping each document in its own tags with metadata such as the source. For document-heavy tasks, it also suggests asking Claude to first quote the passages relevant to the question, then answer based on them.

• Give context about why a rule exists ("the output will be read aloud, so avoid symbols"); Claude uses the reason to generalize.
• Use Projects with project instructions for recurring work, or the system prompt when you use the API.

Gemini: labelled sections, direct wording, instructions after context

Google's prompting guidance for its current Gemini models emphasizes directness. The models are described as less verbose by default, so if you want a chattier or more detailed answer, ask for it explicitly. For complex prompts, Google suggests a clear order: context and source material first, then the main task, then constraints such as format and what to avoid, with the most important constraints near the end so they are not overlooked.

• With large inputs (long documents, codebases, transcripts), put your specific question after the material and anchor it with a phrase like "Based on the document above…".
• If the answer must rely only on the material you provide, say so; otherwise the model may mix in its own knowledge.
• Clearly labelled sections, such as TASK: and CONSTRAINTS:, or consistent Markdown or XML, work well. Pick one style and do not mix them.
• Gems let you save instructions for recurring tasks in the Gemini app; the API has system instructions for the same purpose.

Long documents, system prompts and file inputs

All three apps accept images and documents such as PDFs, and limits on file size and number change regularly, so check each provider's help pages for current figures. The prompting advice for files is the same everywhere: say which file you mean, what you want extracted from it, and whether the model should rely only on that file. With several attachments, refer to them by name ("in contract_v2.pdf, compare clause 4 with clause 4 of contract_v1.pdf"). For images, describe what you want analyzed, for example "read the values in the chart" versus "describe the layout of the slide".

Long prompts also cost more tokens on the API. Paste a draft into the token counter to estimate its size before sending it in bulk.

One task, three prompts

Here is the same task, summarizing customer feedback for a product team, written in the style each provider recommends. The content is identical; only the structure changes.

ChatGPT (Markdown sections)

## Role
You are a product analyst at a project-management software company.

## Task
Analyze the customer feedback below and identify the top 3 recurring
problems.

## Requirements
- For each problem: a one-line summary, how many comments mention it,
  and one representative quote
- Ignore praise unless it contrasts with a problem
- If a comment is ambiguous, leave it out rather than guessing

## Output format
A Markdown table with columns: Problem, Mentions, Quote.

## Feedback
"""
[paste comments]
"""

Claude (XML tags, material first)

<feedback>
[paste comments]
</feedback>

<context>
You are a product analyst at a project-management software company.
The product team will use this summary to plan the next sprint.
</context>

<instructions>
Identify the top 3 recurring problems in the feedback above.
For each: a one-line summary, how many comments mention it, and one
representative quote. Ignore praise unless it contrasts with a problem.
Leave out ambiguous comments rather than guessing.
</instructions>

<output_format>
A Markdown table with columns: Problem, Mentions, Quote.
</output_format>

Gemini (labelled sections, question after context)

CONTEXT:
I am preparing a sprint-planning summary for a product team at a
project-management software company.

FEEDBACK:
[paste comments]

TASK:
Based only on the feedback above, identify the top 3 recurring problems.

CONSTRAINTS:
- For each problem: one-line summary, number of mentions, one quote
- Ignore praise unless it contrasts with a problem
- Leave out ambiguous comments rather than guessing
- Output a Markdown table with columns: Problem, Mentions, Quote

In practice, any of the three prompts would work in any of the three assistants. The adaptations simply follow each provider's published preferences, which can make results a little more consistent, especially with long inputs.

Common mistakes when switching between assistants

Many people use two or three assistants side by side. These are the habits that cause the most friction when you move a prompt from one to another:

• Relying on saved instructions that do not travel. Custom instructions in ChatGPT, project instructions in Claude and Gems in Gemini only apply inside that app. When you copy a prompt elsewhere, paste the relevant parts of those instructions too, or the answer will lose your preferred tone and context.
• Mixing structures in one prompt. Half Markdown headings, half XML tags and a few labels in capitals make it harder to tell where sections begin and end. Convert the whole prompt to one style.
• Expecting the same length. Default verbosity differs between assistants and even between models from the same provider. If length matters, state it ("about 200 words", "five bullets maximum") instead of relying on defaults.
• Burying the question in long material. With long pastes, a question at the very top can be overlooked. Put it after the material, or repeat it there.
• Comparing answers from different contexts. When you test the same prompt in two assistants, use fresh conversations. Earlier messages in a chat change the answer and make the comparison unfair.

Generate the right format automatically

You do not have to remember these conventions. Our model-specific generators build the structure for you, locally in your browser: the ChatGPT prompt generator outputs Markdown sections, the Claude prompt generator wraps every section in XML tags, and the Gemini prompt generator uses labelled sections. If you are writing reusable instructions for an assistant, the system prompt generator applies the same per-model formatting to identity, rules and response style.