Skip to content
Claudexia TeamGUIDE

Gemini prompts: Google's four-part formula and 20 templates

Gemini prompts: Google's four-part formula, 20 templates, app modes and features per Google's help on 17 November 2026, plus the mistakes that waste your time.

A good Gemini prompt answers four questions: who to be, what to do, on what material, and in what form to return the answer. Google puts it this way both in the help for Docs, Sheets and Slides and in its learning materials for Workspace. Below you get the formula, the app's modes per Google's help, 20 ready templates, and the mistakes that cost people time.

The templates are ours. We wrote them from Google's recommendations and did not run them on models, so there are no comparisons or "test results" here. The rules and mode descriptions come from Google pages opened on 17 and 18 November 2026.

What Google says about prompts

We found no single guide on how to write prompts for the Gemini app. The recommendations are scattered across several pages.

  • The Google Docs Editors help singles out four areas: persona (what role to play), task (what to do, with a verb), context (as much background as you can give) and format (the structure of the answer). It also says you do not need all four, and that you should write a prompt the way you speak, in full sentences.
  • The Google Workspace learning center gives five techniques: natural language, clarity and brevity, context, specific terms, and splitting a complex task into separate prompts.
  • The Workspace page on writing prompts adds a technique it calls making Gemini your own prompt editor: start with the phrase "Transform this text into an effective prompt: [your prompt]", check the version it suggests and send that. The same page advises putting the names of your Drive files into the prompt.
  • In a Google blog post of 29 July 2024, the Workspace team writes that in its research the most successful prompts averaged about 21 words, while users' first attempts were shorter than nine words.
  • The Gem help repeats the same four parts for instructions: persona, task, context, format.
  • The Gemini API documentation opens with a caveat: prompt work is iterative, and the recommendations are a starting point.

The four-part formula

PartQuestionExample
PersonaWho to be and how to answer"I am a head of sales"
TaskWhat to do, with a verb"Draft a plan for a client meeting"
ContextBackground, material, goal"The client hesitates over the launch date, the meeting is 30 minutes"
FormatWhat shape to return it in"A table: minutes, topic, question for the client"

Put together, they make one prompt: "I am a head of sales. Draft a plan for a client meeting. The client hesitates over the launch date, the meeting is 30 minutes. Return a table: minutes, topic, question for the client."

To Google's four parts we add two of our own, and this is our recommendation, not a quotation. Constraints: what not to change, what not to invent, how long it should be. Verification: what to do if the data is missing. The phrase "if the answer is not in the material, say so" does not guarantee accuracy, but it gives Gemini a legitimate way to say "I don't know" and makes the gap easier for you to spot.

Gemini modes and features per Google's help

Models and reasoning levels

The "About Gemini models" page lists the app's models, and which ones you get depends on your Google AI plan. The table below is correct as of 18 November 2026 for a personal Google account. For a work or school Google account, Google refers you to the Workspace license.

PlanModels in chatsModels in tasksReasoning levels
No Google AI planAuto or Flash-LiteHydrogenLow, Medium, High
Google AI PlusFlash-Lite, HydrogenHydrogenLow, Medium, High
Google AI ProFlash-Lite, Hydrogen, Pro, ArgonHydrogen, ArgonLow, Medium, High
Google AI UltraFlash-Lite, Hydrogen, Argon, CarbonHydrogen, Argon, CarbonLow, Medium, High

How Google describes the models:

  • Flash-Lite is fast and economical, for short answers and light tasks.
  • 4 Hydrogen balances speed and quality.
  • Pro is for complex prompts, serious code and deep analysis.
  • 4 Argon is called the top-tier model for complex workflows.
  • 4 Carbon is the strongest model, for the hardest long-running tasks.

The reasoning levels work like this. Low gives fast, short answers to simple questions, Medium balances speed and depth for ordinary tasks, and High works through complex prompts in several steps. The higher the level and the stronger the model, the more of your limit it uses. Automatic model selection, when it is on, routes each prompt to the most economical model and level. Without a Google AI plan and with auto selection on, most prompts go to Flash-Lite, and harder ones may go to Hydrogen. We covered how Hydrogen and Carbon differ in our overview of Gemini 4 Hydrogen and Carbon, and the Gemini lineup by task is broken down in our guide to choosing between the Pro and Flash models.

App features

The "Gemini Apps limits and feature availability" page names the features and shows which plans have them. Some rows from its table as of 18 November 2026:

FeatureNo planPlusProUltra
Canvas, Deep Research, Skills, Connected appsyesyesyesyes
Daily briefnoyesyesyes
Scheduled actionsnoyesyesyes
Video generationnoyesyesyes
Tasks in Geminiyesyesyesyes

Google notes that without a Google AI plan, resource-heavy features such as Deep Research may be unavailable during hours of high load.

Google's pages disagree on tasks. In the feature table, Tasks in Gemini is marked on all four plans, and the models page also gives tasks Hydrogen with no plan and on Plus. A separate help article, "Manage your tasks & workflows in Gemini Apps", limits access by country. Without a subscription, tasks exist only in Australia, Canada, India, Japan, South Korea and the United States. With Google AI Pro or Ultra they work everywhere the app does, except the European Economic Area, Nigeria, Switzerland and the United Kingdom. Plus is not named in its lists. Russia is not among the six countries where tasks are open without a subscription. So the "yes" in the table does not guarantee that tasks will work in your country. You also need a personal Google account, an age of 18 or older and History turned on, and for work and school accounts the feature is not yet available.

Here is what the help says about the main features.

  • Canvas. Joint work on a document, an app, slides or code. You select a fragment and ask to change exactly that, and edits save automatically. A finished piece can become an Audio Overview or a quiz.
  • Deep Research. Gemini builds a research plan that you can adjust with the Edit plan button, then compiles a report from many sources. The default source is Google Search, and you can add Gmail and Drive, your own files and NotebookLM notebooks. A report usually takes 5-10 minutes. On Ultra the reports can include charts and interactive simulators, but not when the sources include Gmail, Drive or other Google Workspace services. You must be 18 or older.
  • Gems. Your own assistants with standing instructions. You create them in the web app: Explore Gems, then New Gem, a name, instructions, a preview and Save. The preview does not save anything by itself. The "Use Gemini to re-write instructions" button turns your two lines into a full instruction, and under "Knowledge" you can upload files.
  • Skills. A set of instructions for a recurring task. Google advises writing them like a cheat sheet, not a manual, giving them a lowercase name with hyphens, and adding a description with "Use when…" of up to 1024 characters.
  • Connected apps. In a prompt you type @ and pick an app. On gemini.google.com the History setting must be on for this, and which apps appear depends on your device, country and account type.
  • Attachments. Up to 10 files in one prompt. Video up to 2 GB, other files up to 100 MB. The total video length is up to 5 minutes, and up to an hour with Google AI Pro and Ultra. Audio is up to 10 minutes and up to 3 hours respectively. One code folder or GitHub repository can hold up to 5000 files and 100 MB.

Context window and limits

The context window by plan: 32 thousand tokens without a Google AI plan, 128 thousand on Plus, a million on Pro and Ultra. Google explains that a million tokens is up to 1500 pages of text or 30 thousand lines of code. Limits are counted by compute load, not by the number of prompts, and they refresh every 5 hours until you reach the weekly limit. Deep Research and video generation cost the most, and a short conversation without follow-ups costs very little. The help advises starting a new chat when you change the topic, so Gemini does not have to reread a long history. When you reach the limit on a paid plan, the conversation continues on Flash-Lite.

20 templates

Replace the words in square brackets with your own data. All the templates are ours.

Text and emails

  1. A prompt improver, a technique from the Workspace learning center:
Transform this text into an effective prompt: [your draft prompt].
Add a role, context and an answer format. Return the finished prompt and a list of what you added.
  1. An email built from the four parts:
I am a customer account manager at an online store.
Help me write a reply to a customer: the courier was two days late.
Order [number], the customer writes that [the gist of the complaint].
Format: one paragraph with an apology and three bullet points with ways to resolve it.
  1. A text editor:
Edit the text below. Keep the meaning, names and numbers.
Strip out stiff officialese and cut it to [N] words.
Return the clean version and a list of changes: before, after, reason.

[text]
  1. Meeting notes into tasks:
Here are the notes from a meeting: [text].
Pull out decisions, tasks and open questions.
Return the tasks as a table: task, owner, deadline.
If there is no deadline or owner, write "not specified" and do not invent one.
  1. A product description in three versions:
Write three versions of a description for [product], audience: [who]. Each up to 400 characters, tone: [which].
Use only facts from the text below and add nothing of your own.

[description]

Study and analysis

  1. Explaining a topic:
Explain [topic] to someone at the level of [school, university, beginner in the field].
First two sentences on the core idea, then an example from [field], then three self-check questions.
Give the answers to the questions at the very end in a separate block.
  1. A study plan:
I have an exam in [subject] in [time left] and can study [hours] a day.
Topics: [list]. Make a plan by day.
Format: a table with day, topic, assignment, how to check yourself.
  1. Checking my solution:
Here is my solution to the problem: [solution].
Find the mistakes, but do not rewrite the whole solution.
For each mistake, name the step, the cause and how to fix it.

Files and data

  1. An answer strictly from the document:
A [document] is attached to this prompt. Answer only from its text: [question].
If the answer is not in the document, say so.
Say which section or page you found the answer on.
  1. Comparing two files:
Compare the two attached documents: [first] and [second].
A table: item, what the first says, what the second says, how they differ.
At the end, name the three main discrepancies.
  1. A table and a chart:
A table [file] is attached to this prompt. Describe what is in the columns, and find gaps and odd values.
Suggest three questions this table can answer, and build a chart from [columns].
Show the calculations step by step.
  1. A summary of a recording:
A recording of [a meeting or lecture] is attached to this prompt.
Write a summary in the order of the recording: topic, main points, decisions, unanswered questions.
Do not add anything that is not in the recording.

Code

  1. Reviewing a function:
Here is a function in [language]: [code].
Find bugs and places where it can fail. For each one: the line, what the problem is, the fix.
Do not change the style or the names.
  1. Tests:
Write [framework] tests for the function below.
Cover the normal case, empty input and boundary values.
Do not change the function itself.

[code]
  1. Diagnosing an error:
Here is an error and a piece of code: [log], [code].
List the causes in order of decreasing likelihood and say what to check for each.
If there is not enough data, ask one clarifying question.

Planning and ideas

  1. A project plan:
I am a [role]. Make a launch plan for [project] from [start date] to [end date].
Context: [budget, team, constraints].
Format: a table with week, task, result, main risk.
  1. Ideas with clarification:
I need ten ideas for [topic].
First ask up to three clarifying questions and wait for my answers.
Then number the ideas and give each one a single line of reasoning.

App modes

  1. Deep Research:
Run a research project: [question].
Rely on public sources; do not use my emails or files.
In the report: a short conclusion, three to five facts with sources, what is unknown or disputed.
If the data in the sources contradicts itself, show both versions.
  1. A presentation in Canvas:
In Canvas, make a presentation of [N] slides from the text below.
One slide, one idea, a headline that states the conclusion. Take numbers only from the text.

[text]
  1. Instructions for a Gem:
Persona: you are a technical documentation editor.
Task: bring the texts I send you to one consistent style.
Context: the audience is developers, do not translate terms, do not use colloquial phrases.
Format: first the corrected text, below it a list of changes: "before, after, reason".
If data is missing, ask one clarifying question before editing.

Prompt mistakes

  • A prompt that is too short. In the Workspace team's research, users' first prompts were shorter than nine words, and the successful ones averaged about 21.
  • Several tasks at once. The Workspace learning center and the API documentation advise splitting complex work into separate prompts, and the API docs add chaining them when the answer of one is the input of the next.
  • Vague wording. Google's help says Gemini follows precise instructions, so ambiguity is worth removing.
  • Asking models in the Gemini 2.5 and 3 families to "reason step by step". The API documentation explains that they build their internal reasoning themselves, and spelling out a plan in the answer is usually unnecessary. For hard tasks, a short "think it through carefully" helps, at the price of extra reasoning tokens.
  • Expecting a detailed answer without asking for one. For Gemini 3, the documentation warns that by default the models answer directly and concisely, and a longer or more conversational answer has to be requested.
  • Putting the question at the start of a long context. For Gemini 3, the documentation advises giving all the material first and placing the instruction or question at the very end, for example after the phrase "based on the information above".
  • Fresh facts without search. The API documentation recommends connecting Grounding with Google Search when you need recent or rare information. Both the help and Workspace ask you to check anything important, because Gemini can be wrong.
  • One endless chat for every topic. The help advises a new chat for a new topic, so the limit is not spent on rereading history.

The same rules in the API

The documentation page "Prompt design strategies" advises putting examples in the prompt: Google recommends always including few-shot examples, and if they show the task clearly, part of the instructions can go. Examples should be uniform in format, including tags, whitespace and delimiters. For Gemini 3 the same documentation says to choose one type of delimiter, XML tags or Markdown headings, and not to mix them, and to put the role and hard constraints in the system instruction or at the start of the prompt. There is no separate section on Gemini 4. There is only a note that Gemini 4 and Gemini 3.6 and newer ignore any temperature, top_p and top_k you set, with a warning that in one of the future releases such a request will return a 400 error. It is better not to pass these parameters.

A skeleton following these rules: material first, question last.

<role>
You are a technical documentation editor.
</role>

<constraints>
1. Do not translate terms.
2. Do not add facts that are not in the text.
</constraints>

<context>
[text]
</context>

<task>
Bring the text in the context block to one consistent style and return a list of changes.
</task>

Gemini models can also be called through an OpenAI-compatible gateway. Claudexia is an API gateway with the address https://api.claudexia.tech/v1. In the live price list for 17 November 2026, the Premium channel has gemini-3.1-pro, gemini-3.7-flash, gemini-3.8-flash and gemini-4-argon, and you can find the current list on the models page. A request with a system instruction looks like this, and we did not run it:

curl https://api.claudexia.tech/v1/chat/completions \
  -H "Authorization: Bearer sk_cdx_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-3.8-flash",
    "messages": [
      {"role": "system", "content": "You are a technical documentation editor. Do not translate terms."},
      {"role": "user", "content": "<context>[text]</context><task>Bring the text to one consistent style.</task>"}
    ]
  }'

Registration on claudexia.tech is by login and password or through Google, GitHub and Telegram, the minimum top-up is $1, and you pay in rubles through SBP (Russia's Faster Payments System) or with a Russian bank card, or in cryptocurrency.

Access from Russia

In Google's help, Russia is not in the country list of the Gemini web app or in the country list of Google AI Plus, and in Google's documentation it is not in the list of regions for AI Studio and the Gemini API. Russian is in the web app's language list, though. We give no legal assessments and do not promise that any method is legal or safe for your account. You read and accept the vendor's terms. More in our articles on Gemini in Russia, on getting a Gemini API key in Russia and on why Gemini does not work. The templates in this article suit other chatbots too, and similar breakdowns are in our articles on prompts for ChatGPT and prompts for DeepSeek.

Sources

Pages opened on 17 and 18 November 2026.

FAQ

What is Google's prompt formula?

Four parts: persona, task, context, format. Google writes that using all of them is optional and a few are enough. The same scheme works for Gem instructions.

Can I write prompts to Gemini in Russian?

Russian is in the language list of the Gemini web app in Google's help. Whether English works better for some type of task, Google does not say, and we did not compare the two.

How do I create a Gem?

In the web app, open Explore Gems and click New Gem, name the assistant, write the instructions using the four parts, check them in the preview and click Save. The "Use Gemini to re-write instructions" button expands a short description into a full instruction. Files are added in the Knowledge section.

How is a Gem different from a Skill?

Per Google's help, a Gem is an assistant with standing instructions, and a skill is a set of directions for one recurring task, with steps, format rules and typical mistakes. Gemini reads the full text of a skill only when the skill fits the request.

How many words should a prompt have?

The help gives no exact number. All that is known from the Workspace team's research is that successful prompts averaged about 21 words, while users' first attempts were shorter than nine. That is an average over their observations, not a rule.