How to Write an Effective Gemini Prompt: Guide and Examples

Persona, task, context, format, Gmail and Drive, Deep Research: how to talk to Gemini and make the most of Google without giving it free rein.

8 min read

Milky Way over a desert landscape
Photo: Benjamin Inouye · CC BY 4.0, cropped

Writing to Gemini is like writing to any assistant, with one difference: if you allow it, it can read your emails, your Drive files or your calendar. That's its main appeal, and also the source of a lot of disappointing answers, because a prompt like "summarize my important emails" leaves it to decide what matters to you.

An effective Gemini prompt therefore answers two questions. What you want to get, as anywhere else. And which information it should search.

Google's framework: persona, task, context, format

In its own guides, Google recommends thinking about four elements: the persona (who is speaking or from what point of view), the task (what you want), the context (the situation) and the format (the shape of the answer). Its guidance is flexible: you don't need all four every time, and using a few already helps. Google also advises writing in natural language, in full sentences, as you would to a colleague, rather than in keywords.

Here's what that changes for a bike shop manager who's hiring.

Weak prompt:

Bike salesperson job ad

Gemini returns a standard ad where the team is dynamic and the candidate is passionate. Corrected prompt:

Prompt
You're the hiring manager at a local retail business.

Write a job posting for a full-time salesperson-mechanic position at my bike shop in [CITY]. The team has four people.

Context: we mostly sell electric bikes to individual customers over 50. The person will greet customers in the shop and do small repairs. Saturday work is mandatory, and a half day off is given during the week. Salary: [RANGE].

Format: a clear title, then three short sections (the job, the profile we're looking for, what we offer). 150 words maximum. State the Saturday work and the salary up front: I'd rather candidates who don't agree not apply.

You can find all four elements. The last sentence also gives a reason, which helps Gemini decide what should come first.

How you break up your requests matters too. Google advises splitting complex tasks into several prompts. Don't ask in the same message to analyze your sales, find the causes of a drop and write an action plan. Do it in three messages and check each step: if the starting analysis is wrong, the plan will be too. The method is detailed in the article on prompt chaining.

Tell Gemini where to search when it has access to your apps

In the Gemini app, connected apps let it search Gmail, Calendar, Drive, Docs and other Google services. As of this writing, these connections are managed from the personalization settings, and how far they go depends on your country, your plan and your account type: check in the tool. Google states, for example, that personalization from your apps isn't available in the European Economic Area, Switzerland and the United Kingdom, or for work or school accounts, and that you have to turn it on yourself elsewhere. If you're reading this from France, check what your account offers before counting on it.

When it's enabled, a good prompt narrows the search, as you would with an intern who had access to your mailbox. Specify the source, the period, the sender or the folder, and the goal.

Weak: "Give me an update on my invoices."

Better, for a freelance graphic designer:

Prompt
Search my Gmail emails received since [DATE] for those about pending payments or client reminders, ignoring newsletters and automated notifications.

Present the result in a table: client, subject, date of last message, amount if mentioned, next action to take.

Don't infer any amount: if the email doesn't give one, write "not stated." At the bottom, add the emails you're not sure fall within this scope.

Searching your data can miss a message. A reply that doesn't mention an invoice doesn't prove that the invoice doesn't exist. Use this kind of prompt to clear the ground, never as a final accounting statement.

Connecting apps remains a choice, and you can turn off personalization for a given conversation. Google also says that some conversations are read by human reviewers to improve its services, depending on the activity setting, and advises against entering confidential information you wouldn't want them to see. Read that page of the Gemini help center before running client data through it. Work accounts may be subject to different rules: ask your administrator. For general habits, see what not to paste into a chatbot.

In Docs, Sheets or Gmail, take advantage of what's already open

Gemini in the Workspace side panel isn't quite the same use as the app. It has access to the document you're working in, and Google recommends using the @ symbol to reference other files in your request, where that feature is offered to you. So you don't need to copy the context over: just describe what you want to do with that document.

For a school principal preparing a letter to parents in Google Docs:

Prompt
Using the school council minutes open in this document and the file @[CALENDAR FILE NAME], write a one-page letter to the students' parents.

Keep only what directly concerns them: dates, changes in organization and what they need to do. Write simply, for parents some of whom find English hard to read: short sentences, no acronyms, one idea per paragraph.

If the calendar and the minutes contradict each other, point it out to me after the letter without choosing yourself.

The last line reflects a practice that pays off as soon as several documents are involved: say which one is authoritative, or ask Gemini to flag the conflict. In a spreadsheet, specify the range and the column names rather than "my data," and check any calculation that matters with a formula.

Skills and Gems: write your instructions once

If you re-paste the same instruction every week, you should save it. Google long offered Gems for that: custom versions of Gemini with their own instructions and knowledge files. Its release notes announce that skills, reusable instructions for specific tasks, will replace them and that Gems will be migrated automatically. As of this writing, look in Gemini's personalization menu to see what's available to you. The writing method doesn't change: the instructions work the same for either.

Google advises filling in the same four elements and adding knowledge files if needed. Here are instructions for a real estate agent who wants to write listings consistently:

Prompt
You help a real estate agency in [CITY] write its sale listings.

When I give you the details of a property:
1. Check that there's the floor area, the number of rooms, the floor, the general condition, the heating, the price and the energy diagnostic information. If an item is missing, ask for it before writing.
2. Write a listing of 120 to 150 words: a first sentence that states what sets the property apart, then the description room by room, then the surroundings.
3. Stay factual. Don't use "rare," "exceptional," "love at first sight" or "must see," and don't claim anything about the neighborhood, transit or schools that I haven't told you.
4. At the end, add three titles under 60 characters.

All mandatory legal elements remain my responsibility: just flag the ones you don't see in my information.

Two practical details from the help center: with Gems, the preview doesn't save anything, you have to click Save, and knowledge files added from Drive require Workspace to be connected to Gemini. Once your version is ready, try it on several real cases before relying on it, including one where a piece of information is missing.

For your general preferences, use the instructions for Gemini in the settings instead: your language, your job, your style. The rest goes in a skill or a Gem, as explained in the article on custom instructions.

Prepare Deep Research like an order to a librarian

Deep Research writes a cited report from many sources. Google Search is its default source, and you can add your files or, if Workspace is connected, your Gmail and Drive. The help center says the report takes several minutes. So it's better to ask the question well from the start.

A useful order includes the decision to be made, the geographic and time scope, the type of sources wanted and what you already know. For a nonprofit that wants to open a community grocery:

Prompt
Our nonprofit is considering opening a community grocery in [CITY] within 18 months. We're three volunteers with no paid staff.

Prepare a report to help us decide whether the project is realistic. Cover:
- the legal framework that applies in France and the steps required
- the funding models used by comparable organizations
- the main documented causes of failure
- the national networks that support this type of project

Favor institutional sources and recognized networks, and give the date of the information. Compare my attached files [PROJECTED BUDGET] with what you find. End with the list of what would need to be confirmed with a local contact.

Gemini first creates a research plan, which you can edit before launching the report: take the time to read and correct it, because the number of searches is limited and a poorly aimed report uses one up. Once the report is ready, you can open it in Canvas to work on it, and some plans offer to turn it into visuals. Still, verify the information that will steer your decision. A well-presented, sourced report can still misinterpret its sources, and the article on verification gives you a quick method.

Load the context all at once, but not the task

You often read that you should give Gemini all the context in one go, because it handles large volumes. That's half true, and it squares with Google's advice to split requests up.

Separate what Gemini needs to know from what it needs to do. The context can arrive all at once, in the same message or through attached files. The tasks, on the other hand, are handled one by one as the conversation goes on. A communications officer who submits three annual reports will first ask for the key figures from each, with the page number, then a comparison, then a memo for her management. She'll check at each step.

When several documents contradict each other, say which one prevails: "In case of conflict, the most recent report is authoritative." The result depends a lot on the quality of your files: a scanned PDF or a badly imported table can be misread. A test on a short excerpt doesn't tell you much either: try it on a document of the real size you use.

Now pick a task you do every week in Google Workspace. Write it using Google's four elements, add a line saying where to search, and save the result in a skill or a Gem once it has held up on three real cases.

By tool

Generating images in ChatGPT: what changes in the prompt

In ChatGPT, you ask for an image in conversation and correct it one sentence at a time. How to describe the scene, add text, and iterate without starting over.

3 min read

By tool

Running an AI model on your own computer

What a locally installed AI model makes possible: privacy, the hardware you need, tools to get started, and how to adapt your prompts to a small model.

6 min read

By tool

GitHub Copilot Prompts: Comments, Context, and Chat

How to guide GitHub Copilot: write the comment that comes before the code, choose the chat context, set project rules, and review what it suggests.

4 min read