How to Write Better Prompts for ChatGPT, Claude, and Gemini
Practical, model-agnostic prompting that works on any frontier model. Real before and after examples, reusable templates, and the small per-model quirks worth knowing.
After writing thousands of prompts across every major model, the lesson that sticks is boring: most bad answers trace back to a bad prompt, and switching models rarely rescues a vague request. Learning how to write better prompts is the highest-leverage skill in this whole field, and the good news is that the core moves transfer. A prompt that works on GPT-5.6 usually works on Claude and Gemini too, with a few small adjustments. This guide gives you the patterns, the templates, and the before-and-after examples so you can stop guessing.
The one habit that fixes most prompts
Specificity. That is the whole secret, and it is the thing people skip. When the model produces something vague, generic, or off-target, it is almost always because the prompt was vague, generic, or off-target. The model fills gaps with the most statistically average answer it can find. Your job is to leave fewer gaps.
The widely shared best practices for 2026 are not complicated: give context, be specific, assign a role, show at least one example, ask for step-by-step reasoning on hard tasks, then iterate (see the prompting guidance from MIT Sloan and PromptingGuide.ai). Here is what each one looks like when you actually apply it.
Before and after: turning a weak prompt into a strong one
Take a real, common request. You want help writing a cold outreach email.
Weak prompt:
Write me a sales email.
You will get a templated, lifeless email addressed to nobody about nothing. Now the rewrite.
Strong prompt:
You are a B2B founder writing to another founder, not a marketer. Write a cold email to the head of operations at a 30-person logistics company. We sell route-planning software that cut fuel costs 12 percent for a similar customer. Keep it under 90 words, no buzzwords, one specific question at the end, and a casual subject line. Do not use the words "solution", "leverage", or "reach out".
Same model, completely different result. Notice what changed. The rewrite assigns a role, names the audience, gives a concrete proof point, sets a length, states a format, and bans the words that make AI copy obvious. None of that is clever. It is just refusing to be vague.
Here is a second pair, this time for analysis rather than writing. Say you paste in a spreadsheet of monthly revenue and want insight.
Weak prompt:
What do you think of this data?
You get a generic summary that restates the numbers back to you. The rewrite tells the model what decision you are trying to make.
Strong prompt:
You are a finance analyst. Here is 18 months of monthly revenue. Identify the three months with the largest changes and explain the most likely driver for each. Flag anything that looks like a data error. Then tell me, in two sentences, whether the trend supports hiring a second salesperson next quarter. Show your reasoning before the recommendation.
The weak version asks the model to be impressed by data. The strong version asks it to help you decide something. That framing alone changes the depth of the answer, because now the model knows what the output is for.
If you write for a living, the same principle applies to tone and structure work too, which is why model choice matters alongside prompt quality. We dig into that tradeoff in our guide to the best AI for writing.
Five reusable prompt templates
Templates save you from rebuilding structure every time. Fill in the brackets and adjust to taste. These hold up across ChatGPT, Claude, and Gemini.
Role and constraint: "You are a [role] writing for [audience]. Produce [output] in [format]. Constraints: [length, tone, things to avoid]." Good for any writing or analysis task.
Few-shot by example: "Here are two examples of the style I want: [example A], [example B]. Now do the same for [new input]." One or two examples beat a paragraph of adjectives every time.
Step-by-step reasoning: "Think through this step by step before answering. First list your assumptions, then work the problem, then give a final answer." Use it for math, logic, and multi-part tasks.
Critique and revise: "Draft an answer. Then critique it for [accuracy / clarity / tone]. Then rewrite it fixing those issues." One prompt, three passes, much better output.
Ask before answering: "Before you start, ask me up to three questions if anything is unclear." This single line prevents the model from charging ahead on wrong assumptions.
Treat the first answer as a draft
The single biggest prompt-writing tip that practitioners learn the slow way is to stop expecting the first response to be the final one. Plan on a second pass. Read what came back, find the one thing that is most wrong, and fix that in a follow-up instead of rewriting the whole prompt. "Good, but make the second paragraph shorter and cut the cliche in the opener" gets you further than starting over.
When a response misses badly, the fastest fix is usually to add the missing constraint to the original prompt and rerun it, rather than patching through five rounds of corrections. Every correction round adds length to the conversation, and long threads have their own problems.
Why the AI ignores your instructions
A frequent complaint: "I told it not to do X and it did X anyway." Usually one of three things is happening.
The instruction was buried in the middle of a long prompt. Models pay the most attention to the beginning and end of the input and can lose track of material in the middle, a documented effect researchers call "lost in the middle" (Liu et al., TACL 2024). Put your hard constraints up front or at the very end.
The conversation got long and the early instruction aged out of focus. Quality can degrade as input grows even when the window is far from full. If a long thread starts ignoring your rules, restate them or start fresh.
You phrased a constraint as a negative. "Don't be formal" is weaker than "write casually, like a text to a friend." Lead with the behavior you want, then add the things to avoid.
That middle problem is one reason it pays to understand what a context window actually is. A giant window does not guarantee the model uses all of it well, so a prompt that runs clean in a short chat can drift once the thread grows.
Match the prompt to the reasoning effort
Not every prompt needs a thinking model grinding for thirty seconds. OpenAI exposes a reasoning-effort control with values that can include none, minimal, low, medium, high, and xhigh, and the guidance is to treat medium as the balanced default. Reserve the low settings for fast, simple work and raise the effort only when a hard task clearly benefits.
Reasoning effort helps most on hard math and coding, where accuracy climbs with more thinking. On simpler tasks the gain is small, and pushing the effort too high can even backfire: "overthinking" is a real failure mode where extra reasoning leads a model to abandon a correct answer. Your prompt should signal the depth you need. For a quick rewrite, say so. For a thorny problem, ask explicitly for careful step-by-step work.
If reasoning models are new to you, our explainer on what reasoning models are covers when the extra compute is worth it and when it just slows you down. On MultiChats, paid plans include a session-level reasoning effort control, and our post on reasoning effort control walks through when raising it actually pays off.
Small per-model differences worth knowing
The big patterns are model-agnostic. The fine adjustments are not. Here is a rough field guide to how these families tend to behave rather than a set of hard rules.
Model family | Prompting tip |
|---|---|
ChatGPT (GPT-5.6 class) | Responds well to explicit format and effort cues. State the reasoning depth you want and it adjusts. |
Claude | Rewards structure. Use clear sections or XML-style tags, and it tends to produce more natural prose for writing tasks. |
Gemini | Handles very long inputs and mixed media well thanks to its very large context window. Front-load the key instruction. |
Because the differences are subtle, the practical move is to run the same prompt on two or three models and compare. The model that nails it once is rarely the winner on every task. For a head-to-head on the two most common picks, see ChatGPT vs Claude. And if you are unsure which model fits a given job at all, our guide on how to choose an AI model walks through it.
This is exactly why a multi-model app earns its keep. In MultiChats you can send one prompt, switch the model mid-conversation, and keep the same thread, so testing a prompt across GPT, Claude, and Gemini takes seconds instead of three separate tabs and logins.
A quick prompt-quality checklist
Before you hit send on anything important, run through this.
Did I give the model a role and an audience?
Did I state the format and length I want?
Did I include an example if style matters?
Are my hard constraints at the start or the end, not buried?
For a hard task, did I ask for step-by-step reasoning?
Frequently asked questions
How do I write a good AI prompt?
Be specific and give context. Assign a role, name your audience, state the format and length, and add one example if style matters. For anything difficult, ask the model to reason step by step. Then iterate: read the first answer, see what is off, and tighten the prompt. Most of the improvement comes from removing vagueness.
Do the same prompts work on ChatGPT and Claude?
Mostly, yes. The core moves (role, context, examples, format, step-by-step reasoning) transfer across ChatGPT, Claude, and Gemini. You will see small differences in tone and formatting, and Claude in particular tends to respond well to clearly structured prompts with sections or tags. The fastest way to find out is to run the same prompt on both and compare.
What is prompt engineering in 2026?
It is the practice of writing inputs that get reliable, high-quality outputs from a model. In 2026 the secret-magic-phrase era is mostly over. What works now is clear communication: context, specificity, examples, structure, and choosing the right reasoning effort for the task. Models are far more capable than they were a couple of years ago, so the payoff comes from being precise. Tricks and jailbreak phrasings matter much less than they used to.
How long should a prompt be?
As long as it needs to be to remove ambiguity, and no longer. A two-line prompt is fine for a simple task. A detailed task may need a paragraph of context plus an example. Avoid padding: extra words can bury your real instructions, and long inputs can suffer from the lost-in-the-middle effect. Put the most important constraints at the start or the end.
The short version
Better prompts are specific prompts. Give context, assign a role, show an example, state the format, and ask for reasoning when the task is hard. Those habits work on any frontier model, so the skill you build is portable. The fastest way to get good is to test the same prompt on a few models and watch how they differ.
Want to test your prompts on GPT-5.6, Claude, and Gemini in one thread without juggling subscriptions? Try MultiChats with one plan for 25+ models.