Google Gemini 101: from your first chat to the API

If you have an Android phone, a Gmail account or have searched Google recently, you have probably already met Gemini. It is Google’s AI — a chat assistant you can talk to, point your camera at, and ask to research the web for you — and it is quietly being built into almost everything Google makes.

This guide takes you from your first question to your first program. We begin with where Gemini came from and what makes its design different, pick the right model, then go step by step through the app on the web and on your phone, its most useful features, Gmail and Docs, and finally the developer side: Google AI Studio, the Gemini API and the Gemini CLI.

What is Gemini?

The word Gemini means two things. It is Google’s family of AI models, built by its research lab Google DeepMind. And it is the assistant you use — the Gemini app — along with the AI features inside Android, Chrome, Gmail, Docs and Search that run on those models.

Its biggest strengths come from Google itself: it can search the web with Google Search as it answers, it works inside the apps billions of people already use, and it was designed from the start to understand images, audio and video, not only text.

Google’s AI story

Google has a special place in this history: the Transformer, the design at the heart of ChatGPT, Claude and Gemini alike, was invented by Google researchers in 2017. Yet it was OpenAI that put a chatbot in front of the public first. Gemini is the result of Google’s push to catch up — and, in areas like multimodality and long context, to lead.

How Gemini is built

Like every modern chatbot, Gemini is a large language model based on the Transformer: it turns its input into tokens and predicts, one token at a time, what should come next. (Our first article, Welcome to the world of AI, explains that from scratch.) Two design choices set Gemini apart.

1. Natively multimodal

Early AI assistants handled pictures and sound by bolting on extra systems: one converts speech to text, another writes a caption for an image, and only then does the language model see anything. Gemini was trained from the beginning on text, images, audio and video together, so all of them become tokens in the same stream. That is why it can notice the tone of a voice, read a chart inside a photo, or follow what happens across a video.

2. Mixture of experts

Google’s technical reports describe Gemini models using a mixture-of-experts (MoE) design. Instead of one giant block of neurons that all fire for every word, the model contains many smaller “expert” networks and a router that sends each token to just a few of them. The model can be enormous in total while each step stays fast and affordable.

The model family

Gemini comes in tiers. Flash-Lite is the cheapest and quickest; Flash balances speed and intelligence and is Google’s recommended starting point; Pro reasons the most deeply. Around them sit specialists for live voice, speech, images (the famous “Nano Banana”), video, music and embeddings.

Newer Gemini models are thinking models: before answering, they can reason through a problem internally. In the app this appears as a choice between a fast mode and a thinking mode; in the API, developers can control how much the model thinks, trading depth against speed and cost.

Getting started, step by step

1. Open Gemini. Go to gemini.google.com on a computer, or install the Gemini app on iPhone. On most recent Android phones, Gemini is already built in.
2. Sign in with your Google account. The same account that runs your Gmail. Work and school accounts may need an administrator to switch Gemini on.
3. Learn the screen. It has only a few parts:

4. Ask for something real — a trip plan, an explanation, a draft — and keep going with follow-ups.
5. Add files and photos with the ＋ button: a bill to explain, a PDF to summarise, a screenshot of an error.
6. Use “double-check”. Under an answer, the menu offers a check against Google Search that highlights statements with supporting or conflicting sources.

Gems, Deep Research, Canvas and Live

Gems: your own custom Geminis

A Gem is a saved version of Gemini with its own instructions and, optionally, its own files. Write the instructions once — who it is, what it does, how it answers — and open it whenever you need it: an interview coach, a recipe planner, a writing editor in your house style.

Deep Research: a report while you make tea

For questions that need many sources, choose Deep Research. Gemini drafts a research plan you can edit, then browses and reads for several minutes, and returns a structured report with its sources, which you can export to Google Docs.

Canvas and Live

Canvas opens a workspace beside the chat for longer documents and code, which you and Gemini edit together; it can also turn material into web pages, quizzes and audio overviews. Gemini Live is a spoken conversation: tap Live, talk naturally, interrupt whenever you like, and on a phone share your camera or screen so Gemini can see what you see.

Gemini on your phone

On Android, Gemini is the phone’s built-in assistant. Press and hold the power button (or say “Hey Google”, if enabled) to bring it up over whatever you are doing, and ask about what is on screen. On iPhone, install the Gemini app from the App Store. On both, Live with the camera is the stand-out feature: fix a bike, identify a plant, translate a sign, or get help with homework on paper.

Gemini in Gmail, Docs and Drive

On eligible plans, a Gemini button appears inside Google Workspace apps and opens a side panel that can see the email, document or folder you are working on. It summarises long threads, drafts replies, turns rough notes into a document, builds formulas and tables in Sheets, and finds files in Drive by what is in them.

For deep work across your own sources, Google also offers NotebookLM: upload papers, notes or videos, and ask questions answered only from those sources, with citations — even as a podcast-style audio overview.

Prompting Gemini well

The prompting habits from ChatGPT 101 — role, task, context, format, examples, limits — work just as well here. A few tips make the most of Gemini’s particular strengths:

• Show, don’t describe. Gemini reads images, audio and video natively. Instead of typing out a table from a photo, just attach the photo.
• Refer to parts precisely. With long material, point to where you mean: “in the third section”, “at 12:40 in the video”, “the chart on page 7”.
• Ask for sources when facts matter. “Use Google Search and link your sources” nudges Gemini to ground its answer.
• Use the right tool. A quick question goes in chat; a big one goes to Deep Research; a long document goes in Canvas; a repeated job becomes a Gem.
• Reach into your apps by name. Mentioning Gmail, Drive or Calendar tells Gemini where to look.

Plans and privacy

The Gemini app is free with a Google account, and the free tier is generous: Deep Research, Canvas, Gems, Live and image creation are all included. Paid Google AI plans raise the limits, unlock more of the top models and video tools, add Gemini inside Gmail and Docs, and bundle extra Google storage.

Privacy checklist

• Open Gemini Apps Activity in the app’s settings to see what is saved, set auto-delete, or turn saving off. Some conversations may be reviewed by people to improve Google’s services, so read the notice there.
• Use temporary chats for anything you would rather not keep.
• Don’t share secrets: passwords, card and bank numbers, Aadhaar or passport numbers.
• Work and school accounts follow your organisation’s data terms, which are usually stricter.

Long context and grounding

Gemini made long context famous: Gemini 1.5 Pro was one of the first models to read about a million tokens in one go, and large context windows remain a Gemini strength. That changes what is possible — you can hand over a whole book, an hour-long lecture video or a sizeable codebase and ask questions across all of it.

Long context is not free: bigger inputs cost more and take longer. Two techniques help. Context caching stores a large, unchanging input — say, a 500-page manual — so repeated questions reuse it at lower cost. And retrieval (RAG) searches your material first and sends only the relevant chunks, which scales to libraries far bigger than any context window.

Grounding with Google Search

A model’s built-in knowledge stops at its training date. Grounding connects Gemini to Google Search so it can look things up while answering and cite what it found. In the API it is a built-in tool you switch on; Gemini decides when a question actually needs a search.

Google AI Studio: the developer’s playground

Google AI Studio (aistudio.google.com) is where developers try Gemini before writing code. Sign in with a Google account, pick a model, write system instructions, attach images or files, switch on tools such as grounding or structured output, and see the result immediately. When it works, Get code produces matching code in several languages, and Get API key creates the key you need.

For company use, the same models are offered on Vertex AI in Google Cloud, with enterprise security, data residency and billing through a Cloud account.

The Gemini API

Install Google’s SDK with pip install -U google-genai, create a key in AI Studio, and set it as GEMINI_API_KEY. Google’s current documentation centres on the Interactions API: each call is an interaction with a model, an input and optional instructions.

from google import genai

client = genai.Client()  # reads your GEMINI_API_KEY

interaction = client.interactions.create(
    model="gemini-3.8-flash",
    system_instruction="You are a patient maths tutor for Class 10 students in India.",
    input="Explain why (a + b)² = a² + 2ab + b² using a picture I can draw.",
)

print(interaction.output_text)

To hold a conversation, pass the previous interaction’s ID and Gemini continues from there — no need to resend the whole history yourself:

first = client.interactions.create(
    model="gemini-3.8-flash",
    input="I have ₹20,000 to invest every month for 10 years.",
)

follow_up = client.interactions.create(
    model="gemini-3.8-flash",
    input="Show me what that becomes at 8% a year, as a table.",
    previous_interaction_id=first.id,   # continue the same conversation
)
print(follow_up.output_text)

Structured output

For software, ask for JSON that matches a schema. With Pydantic you define the shape once and get validated objects back:

from pydantic import BaseModel

class Expense(BaseModel):
    merchant: str
    amount_inr: float
    category: str

interaction = client.interactions.create(
    model="gemini-3.8-flash",
    input="Paid 1,240 at Big Bazaar for groceries and snacks on Friday.",
    response_format={
        "type": "text",
        "mime_type": "application/json",
        "schema": Expense.model_json_schema(),
    },
)

expense = Expense.model_validate_json(interaction.output_text)
print(expense)  # merchant='Big Bazaar' amount_inr=1240.0 category='Groceries'

And the same basic call over plain HTTP:

curl -X POST "https://generativelanguage.googleapis.com/v1beta/interactions" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-3.8-flash",
    "input": "Give me three names for a momo stall in Shillong."
  }'

What else the API can do

• Multimodal input: send images, audio, video and PDFs alongside text.
• Function calling: describe your own functions; Gemini asks you to run them and uses the results.
• Built-in tools: Google Search grounding, code execution and URL reading.
• Live API: real-time, two-way voice and video for conversational apps.
• Generation: images with Nano Banana, video with Veo and Omni, speech, and music.
• Embeddings: turn text and media into vectors for semantic search.

Gemini CLI: an agent in your terminal

Gemini CLI is Google’s open-source AI agent for the command line. It reads and edits files, runs shell commands with your permission, searches the web with Google Search, and connects to other tools through the Model Context Protocol (MCP). Signing in with a personal Google account includes a free allowance — at the time of writing, 60 requests a minute and 1,000 a day.

# run without installing
npx @google/gemini-cli

# or install it
npm install -g @google/gemini-cli
brew install gemini-cli        # macOS / Linux

gemini                          # start, then sign in with Google

$ gemini
✦ Gemini CLI  · signed in with Google · GEMINI.md loaded

> Rename every photo in this folder by the date it was taken, like 2026-09-14_001.jpg

✓ ReadFolder . — 214 .jpg files
✓ Shell exiftool -DateTimeOriginal IMG_0001.jpg

I’ll write a small script that reads each photo’s EXIF date and renames it.
3 photos have no date — I’ll leave those unchanged and list them.

? Allow execution of: python rename_by_date.py --dry-run
  › Yes, allow once   Always allow   No

✓ Shell dry run: 211 renames, 3 skipped — no clashes

Looks right. Run it for real?

Put a GEMINI.md file in your project to give the CLI lasting context — how to build and test, conventions, things to avoid — and it will read it every session. The same ideas power Google’s Antigravity, an agent-first development environment where several agents can work on tasks in parallel.

Your 30-day path

Small steps, one week at a time, from first chat to first program.

Quick quiz

Tap a question to check your answer.

That is Gemini, from holding the power button to your first line of code. Start with the door closest to you — probably the phone in your pocket — and build from there. ✦

Gemini, Google AI Studio, Gemini CLI, NotebookLM and Android are products of Google; this article is independent and not affiliated with or endorsed by Google. Model IDs and API code follow the Gemini API documentation, CLI details follow the Gemini CLI repository, and plan features follow Google’s subscriptions page, all as of 24 September 2026. Interface screens are illustrations and sample outputs are examples. Spotted something out of date? Tell us.