If you are using ChatGPT, Claude, and Gemini in 2026, you already know they do not think the same way. But many people still copy and paste the exact same prompt into all three. This is a mistake. Each model has its own preferred syntax, structure, and pacing. If you want the best results, you need to speak their language.
In this guide, I will break down exactly how to format your prompts for the big three AI models. I built Kosmo because I got tired of rewriting my prompts for every new model update. Now, I want to show you the rules we use under the hood to get the best out of ChatGPT (GPT-5.5), Claude (Opus 4.8), and Gemini (3.1 Pro).
Which AI is best for your use case?
Before we look at the code, let us look at the landscape. As of mid-2026, the big three are priced almost identically for consumers, but they serve different needs. OpenAI, Anthropic, and Google all charge around $20 a month for their flagship tiers. All of them now offer massive context windows, usually crossing the 1 million token mark.
| Feature | ChatGPT Plus | Claude Pro | Google AI Pro |
|---|---|---|---|
| Flagship Model | GPT-5.5 / 5.6 | Fable 5 / Sonnet 5 | Gemini 3.1 Pro |
| Context Window | 1 Million+ Tokens | 1 Million+ Tokens | 1 to 2 Million Tokens |
| Prompting Style | Markdown & Modular | XML Tags | Direct & Few-Shot |
| Monthly Price | $20.00 | $20.00 | $19.99 |
How do you write the best prompt for ChatGPT?
ChatGPT thrives on markdown and system-first design. OpenAI trained the GPT-5 family to act like an agent. It wants a clear role, explicit boundaries, and modular instructions.
The best practice for ChatGPT is the Bento-Box pattern. You separate your instructions, context, and data using clear markdown headers. You also want to tell it to think step by step if the task is complex.
Here is how you format a task to extract action items from a transcript for ChatGPT:
# Role
You are an executive assistant.
# Task
Read the following transcript. Extract all action items.
Think step-by-step before finalizing the list.
# Constraints
- Format as a markdown checklist.
- Include the name of the owner for each item.
# Transcript
[Insert text here]What is the right prompt style for Claude?
Claude loves XML tags. Anthropic tuned models like Opus 4.8 and Fable 5 to look for semantic boundaries. If you use XML tags like <instructions> or <context>, Claude parses the prompt almost flawlessly.
You should treat a Claude prompt like a legal contract. Give it clear rules inside these tags. You do not need to tell Claude to think step by step. It already does that internally.
Here is the exact same task, rewritten for Claude:
<role>You are an executive assistant.</role>
<instructions>
Read the transcript and extract all action items.
Format the output as a bulleted list.
Include the owner for each item.
</instructions>
<transcript>
[Insert text here]
</transcript>How should you prompt Gemini?
Gemini wants you to get straight to the point. Google optimized Gemini 3.1 Pro for speed, high throughput, and direct commands. The "less is more" rule applies here. Cut the conversational fluff.
Gemini also likes structural tags to separate data, but it relies heavily on few-shot examples. If you want a specific output, show Gemini an example.
Here is the Gemini version of our prompt:
Task: Extract action items from the transcript.
Rules: Format as a checklist. Note the owner. No introductory text.
Example Output:
- [ ] Task 1 (Owner: Sarah)
- [ ] Task 2 (Owner: John)
Transcript:
[Insert text here]How Kosmo handles this automatically
I built Kosmo because keeping up with these rules is exhausting. You should not have to memorize whether a model prefers XML tags or markdown headers.
With Kosmo, you write your intent once in plain English. You say, "Extract action items from this transcript and format as a checklist." Kosmo acts as an AI prompt compiler. It detects which model you are targeting and automatically injects the right XML tags for Claude, the Bento-Box markdown for ChatGPT, or the direct examples for Gemini. So you get the optimal output every time, without the manual formatting.
Frequently asked questions
Does the context window size matter for prompting?
Yes. All three major models now support at least 1 million tokens. But just because you can paste a whole book does not mean you should. A massive context window still requires clear structure. Always put your instructions at the very top or the very bottom of a long prompt to prevent the model from forgetting the goal.
Do I still need to assign a role to the AI in 2026?
Assigning a role remains highly effective for ChatGPT and Gemini. It helps the model calibrate its vocabulary and tone. Claude relies a bit less on roles and more on explicit instruction tags, but a clear <role> tag still helps frame the context.
Which model is the cheapest to run via API?
While consumer plans sit at around $20 per month across the board, API costs vary wildly. Google's Gemini 3.1 Pro is often positioned as the value choice for high-volume tasks, while OpenAI and Anthropic charge a premium for their flagship reasoning models. Always check the official pricing pages before scaling an app.