A Simple Prompt Framework That Works in ChatGPT, Gemini and Claude

Most disappointing AI answers start with a disappointing question. "Write a marketing email" gets you something generic because the assistant has to guess everything that matters: who it is for, what you are selling, how long it should be and what tone fits your brand.

The good news is that you do not need tricks or secret phrases. Google, Anthropic and OpenAI all publish guidance on how to write prompts for their assistants, and the core advice overlaps heavily. This article distills it into a framework you can remember, then shows how to apply it with examples.

The four parts of a good prompt

Google's prompting guide for Gemini in Workspace organizes prompts around four elements: persona, task, context and format. Anthropic's documentation for Claude makes the same points in different words: be clear and direct, add context, use examples and specify the output you want. Together they give a reliable checklist.

1. Persona: who should the assistant be?

A single sentence that sets a role changes the vocabulary, depth and tone of an answer. "You are an experienced accountant explaining things to a first-time freelancer" produces a very different reply from "You are a tax attorney writing for other attorneys". Anthropic's guide notes that even a one-sentence role makes a difference.

2. Task: what exactly do you want done?

Use a clear verb: summarize, draft, compare, rewrite, critique, list. If the order of steps matters, number them. Anthropic suggests a useful test: imagine handing your prompt to a capable colleague who knows nothing about your project. If they would have to ask follow-up questions, the assistant will have to guess, and guesses produce generic output.

3. Context: what does it need to know?

This is the part people skip, and it matters most. Include the audience, the goal, constraints, and any material the answer should be based on. Google's guide tells users to provide as much context as possible, including pointing Gemini at existing documents in Drive.

Context also includes why. Anthropic gives a memorable example: instead of "never use ellipses", explain that the response will be read aloud by a text-to-speech engine that cannot pronounce them. When the model understands the reason, it applies the rule more sensibly, including in cases you did not think to mention.

4. Format: what should the answer look like?

Say how long, how structured and in what shape: "three bullet points", "a table with columns for price, pros and cons", "under 120 words", "a subject line plus a short email". Anthropic recommends describing what you want rather than what you do not want. "Write in flowing paragraphs" works better than "do not use bullet points".

You do not need all four in every prompt. Google's guide is explicit that using a few of them already helps. But when an answer disappoints, the missing piece is almost always one of these four.

Before and after: three examples

An email

Before: "Write an email asking for a deadline extension."

After: "You are a project coordinator at a small design agency. Draft a short, polite email to a client asking to move the delivery of their website mockups from Friday to next Wednesday. The reason: their brand guidelines arrived three days late. Keep it under 120 words, take responsibility for the new date, and end with a clear yes/no question."

The second version adds a persona, a precise task, the real reason (context) and a length and ending (format). The result needs far less editing.

A summary

Before: "Summarize this report."

After: "Summarize the report below for our sales team, who have two minutes to read it before a client call. Give five bullet points covering only findings that affect pricing or delivery times, then one sentence on what we should tell the client. Report: [paste report]"

A decision

Before: "Which laptop should I buy?"

After: "I'm choosing a laptop for video editing in DaVinci Resolve, mostly 4K footage from a phone. Budget is around $1,500 and I travel weekly, so weight matters. Compare three suitable options in a table (price, weight, GPU, battery life), then recommend one and explain the trade-off in two sentences. Tell me which specs I should double-check on the manufacturer's site, since your information may be out of date."

Note the last sentence. Assistants can be wrong about prices and specifications, especially for recent products. Asking the model to flag what to verify is a habit worth building.

Three techniques that add a lot

Show an example of what you want

If you need a particular style or structure, paste one or two examples. Anthropic calls examples one of the most reliable ways to steer format and tone, and recommends three to five varied examples for best results. Make them relevant to your real task and different enough from each other that the model does not copy accidental details.

Separate instructions from material

When a prompt mixes instructions with a long document, label the parts. Anthropic's guide suggests wrapping each part in descriptive tags such as <instructions>, <context> and <document>. Plain headings like "INSTRUCTIONS:" and "TEXT TO EDIT:" work in any assistant. The point is that the model never confuses the text you want edited with the instructions about how to edit it.

Put long material first and the question last

For long inputs, such as a contract, a transcript or a set of reports, Anthropic recommends placing the documents at the top and your question at the end, noting that in its tests this improved answer quality by up to 30 percent on complex, multi-document inputs.

Treat it as a conversation

The first answer is a draft. Google's guide puts heavy emphasis on refining with follow-up prompts, and this is where most of the quality comes from. Useful follow-ups:

• "Make it more direct and cut it by a third."
• "Which of these points is weakest? Replace it."
• "Rewrite this for someone who has never heard of our product."
• "What did you assume that I didn't tell you?"

That last question is especially useful. It surfaces the guesses the assistant made, which are usually exactly the context you forgot to give.

It also helps to split a big job into steps rather than asking for everything at once. Ask for the outline, approve it, then ask for the draft. Each step is easier to check, and errors do not compound.

A template you can copy

You are [persona].
Task: [verb + exactly what you want].
Context: [audience, goal, constraints, why it matters].
Material: [paste text, or say "none"].
Format: [length, structure, tone].
Before answering, list anything you would need to assume.

One rule that never changes

Better prompts produce better drafts, not guaranteed facts. Google's guide explicitly tells users to make sure responses are accurate before using them. Check names, numbers, dates and quotes before you send or publish anything, and be especially careful with medical, legal and financial topics, where a confident wrong answer can cause real harm.

Sources

• Google Workspace: Writing effective AI prompts
• Anthropic: Prompting best practices
• Anthropic: Prompt engineering overview