ChatGPT, Claude and Gemini all read plain English, and a clear request gets a decent answer from any of them. The difference shows up on harder jobs: long pasted documents, strict output formats, several steps that must happen in order. Each assistant follows some structures more reliably than others, and writing for the one you use saves you the second and third attempt.
This guide takes one real task and writes it three ways, then covers the parts every good prompt shares. If you would rather not do it by hand, the ChatGPT prompt generator, Claude prompt generator and Gemini prompt generator write the prompt in each structure from a one-line idea.
The five parts every prompt needs
Before the differences, the common ground. A prompt that works on the first try usually answers five questions:
- Who is answering. One line of role ("You are a support lead at a software company") sets vocabulary and judgement. Skip long biographies; one line does the job.
- What the task is. One sentence, with a verb: summarise, draft, compare, rewrite, plan.
- What the assistant needs to know. The facts it cannot guess: your audience, your product, the document itself, what you tried already.
- The rules. Length, tone, things to avoid, what to do when information is missing.
- The shape of the answer. A table, five bullets, an email with a subject line, JSON with named fields. Say it exactly.
Most weak prompts skip the third and fifth parts. The assistant then fills the gaps with generic assumptions, and you get a generic answer.
A habit worth keeping: when the assistant needs a fact you have not given, leave a placeholder in square brackets, such as [PRODUCT NAME] or [DEADLINE]. You fill it in before sending, and the assistant never invents it.
The example task
Here is the rough idea we will turn into three prompts:
Turn my messy meeting notes into a short decision log my team can read in a minute.
On its own, this gets you a tidy summary of whatever the assistant thinks matters. With structure, it gets you the decisions, the owners and the open questions, in the format you want.
Writing for ChatGPT
For ChatGPT we use a short role line followed by clearly labelled sections. Plain labels with simple dashes are enough; markdown headings inside the prompt add nothing.
Why this shape works: the labels separate your instructions from your material, and the Output line removes any doubt about format. If the answer drifts, the fix is usually a sharper Requirements line, not more adjectives.
Writing for Claude
Claude responds well to XML-style tags that wrap each part of the prompt. Anthropic's own prompting guidance recommends tags for exactly this reason: when you paste a long document, the tags mark where your material starts and ends, so Claude does not treat a sentence in your notes as an instruction.
Numbered steps inside the instructions tag help when the order matters. The tag names themselves are not magic; any clear, consistent names work. What matters is that each part sits in its own container.
Writing for Gemini
For Gemini we put the task in the opening sentence, then short and literal sections. A closing quality check line gives it something concrete to verify before it answers.
Short sentences matter here. Long sentences with several clauses leave more room for the model to pick the wrong one as the main instruction.
What changes between the three, side by side
| ChatGPT | Claude | Gemini | |
|---|---|---|---|
| Opening | One-line role | One-line role | The task itself |
| Sections | Plain labels | XML-style tags | Short labelled parts |
| Steps | Optional | Numbered, inside the instructions tag | Numbered |
| Best for | Everyday drafting and quick formats | Long pasted documents and strict rules | Literal tasks with a clear check |
None of this is a hard rule. All three assistants will read any of the prompts above and do a reasonable job. The structure tips the odds toward the answer you wanted.
Five mistakes that cost a second try
- Asking for "the best" without saying best for whom. Name the audience and what they need to do with the answer.
- Hiding the format at the end of a long paragraph. Put the output format on its own line.
- Mixing material and instructions. Paste documents inside a labelled block or tags, never in the middle of your request.
- Stacking tone words. "Friendly, professional, engaging, concise" cancels itself out. Pick one tone and one length.
- Letting the assistant fill in facts. Use placeholders for anything you have not given it, and say what to do when information is missing.
Starting points from the library
If you want a finished prompt to adapt rather than a blank page, the text prompt library groups prompts by job: writing, business, marketing and study. The meeting notes to decision log prompt is a longer version of the example in this post.
Let the generator write it
Typing structure by hand is fine for a prompt you will reuse. For a one-off, describe the idea in a sentence on the AI prompt generator, pick ChatGPT, Claude, Gemini, DeepSeek or Grok as the target, and choose the tone and answer length. The prompt comes back in that assistant's structure, with placeholders where it needs facts from you. You can then ask for it shorter, more detailed, more formal or with an example of the expected output, or switch the target and get the same prompt rewritten for another assistant.
