GUIDE

How to prompt ChatGPT, Claude and Gemini

OpenAI, Anthropic and Google each publish advice on prompting their own models. Most of it agrees: say what you want, give the context, and be specific about the result. The differences are in the details, such as how many examples to include and where long material goes, and they are worth knowing if you use more than one assistant.

We read each vendor’s current guides and summarized what they recommend, with a link to every source. Vendors revise these pages often; this summary was checked on September 24, 2026.

Create my prompt Describe your goal, choose where you will use it, and get a finished prompt.

Updated September 24, 2026

Where the advice agrees

Be clear about the result you want. OpenAI’s ChatGPT guide says to start with the result, not a detailed list of steps. Anthropic suggests a test: if a colleague with little context would be confused by your prompt, the model will be too.
Give the context the model cannot know: who the answer is for, what already exists and how it will be used.
Separate instructions from material. OpenAI and Anthropic both suggest delimiters such as XML tags or Markdown headings to show where each part begins and ends.
Split big jobs. Google suggests one prompt per instruction, or a chain of prompts for work with several steps.
Say what you want back: the format, the length and the level of detail.

Where it differs

What each vendor recommends, by topic
Topic	OpenAI (ChatGPT)	Anthropic (Claude)	Google (Gemini)
Examples	For reasoning models, try without examples first and add them only if needed.	Use 3–5 examples for best results, each in its own tag.	Always include a few examples, all in the same format.
Long material	Mark where each document begins and ends with delimiters.	Put long documents first and the question last: up to 30% better responses in Anthropic’s tests.	Put the context first and your question at the very end.
Step-by-step instructions	Unnecessary for reasoning models, which reason internally.	A general “think thoroughly” often beats a hand-written step-by-step plan.	Gemini 3 has a thinking setting that replaces step-by-step prompting.
Emphasis	Avoid contradictory or vague instructions; GPT-5 spends effort reconciling them.	Dial back aggressive wording such as “CRITICAL: You MUST”.	Be precise and direct, without overly persuasive language.
A starting framework	Goal, context, output and boundaries.	Clear, direct instructions, with the reason behind each one.	Persona, task, context and format.

Prompting ChatGPT

For bigger tasks, OpenAI’s ChatGPT guide suggests four parts, used only when they help: the goal, the context, the output you need, and the boundaries, meaning what must stay unchanged.
Start with the result you want rather than a detailed list of steps, then refine with follow-ups.
With OpenAI’s reasoning models, give specific criteria for a successful answer instead of asking them to think step by step, and try without examples first.
Remove contradictions. OpenAI’s GPT-5 guide says contradictory or vague instructions can do more damage, because the model spends reasoning effort trying to reconcile them.
Use delimiters, such as Markdown headings or XML tags, to mark the distinct parts of a long input.

Prompting Claude

Anthropic describes Claude as a brilliant but new employee who lacks context on your norms: the more precisely you explain what you want, the better the result.
Explain why an instruction matters, not just what it is. Anthropic says the motivation behind an instruction helps Claude give more targeted responses.
Wrap each kind of content, such as instructions, documents and examples, in its own XML tag, and give 3–5 examples for best results.
With long inputs, roughly 20,000 tokens or more, put the documents first and your question at the end.
Say what to do rather than what not to do, write the prompt in the style you want back, and dial back all-caps urgency, which newer models can over-react to.
A role in the system prompt, even a single sentence, focuses Claude’s behavior and tone.

Prompting Gemini

Google recommends always including a few examples, and says prompts without them are likely to be less effective. Keep every example in the same format.
Put the essentials, such as the persona, constraints and output format, in the system instruction or at the start of the prompt.
With a lot of context, put your specific question at the very end.
Keep it concise. Google says Gemini 3 may over-analyze verbose or heavily engineered prompts written for older models, and that it is less verbose by default, so ask for more detail when you want it.
Google’s Workspace guide frames a prompt as persona, task, context and format, and reports that the most fruitful prompts it sees average around 21 words with relevant context, while the prompts people try are usually under nine words.

Coding assistants: Cursor and Codex

Both read standing instructions from files. Cursor supports project rules in .cursor/rules and a simpler AGENTS.md file; Codex reads AGENTS.md files from the repository root down to your working folder, and files closer to your folder override earlier guidance.
OpenAI’s advice for Codex: name the behavior you want, point to the relevant code or reproduction steps, keep the important constraints, and say how to verify the change.
Cursor suggests planning before building and, when a result is off, reverting and refining the plan rather than patching it with follow-up prompts.
Keep rule files lean. Cursor advises keeping rules under 500 lines and adding a rule only when the agent keeps making the same mistake.

More on writing briefs for coding tools: coding prompts.

What this means for your prompts

The substance travels. A clear goal, the context, the must-haves, the limits, the format and a self-check work in every assistant.
The packaging changes. More examples for Gemini and Claude, fewer for OpenAI’s reasoning models; long material before the question for Claude and Gemini.
Old tricks age. Lines like “think step by step” helped older models, but OpenAI and Google now say their reasoning models do this work themselves.

Mint Prompts writes the substance for you: describe your goal, say where you will use the prompt, and copy the finished prompt into that tool. To write one yourself, start with how to write a good AI prompt.

Questions

Does a prompt written for ChatGPT work in Claude or Gemini?

Usually, yes. The substance of a good prompt, its goal, context and limits, works everywhere. What varies is the packaging, such as how many examples to include and where long material goes.

Should I use XML tags in my prompts?

They help when a prompt mixes several kinds of material, such as instructions, documents and examples. Anthropic recommends them for Claude, and OpenAI lists XML tags among useful delimiters. For a short request, plain sentences are fine.

Do I still need to write “think step by step”?

Less than before. OpenAI says it is unnecessary for its reasoning models, and Google offers a thinking setting for Gemini 3 instead. A clear goal and a way to check the answer do more.

How often does this advice change?

Often. Each vendor revises its guides as new models ship, so follow the source links for the current version. This page was last checked on September 24, 2026.