Your first conversation locale

#What you'll be able to do in this tutorial

You will ask for a summary, have an email written, request a small script, test a question outside the model's scope, then correct an answer that is heading in the wrong direction. These are the actions that make you self-sufficient. The guide assumes Ollama is already installed; if not, start with your system's installation guide. To go further on how to phrase an instruction, the guide to prompting basics takes over.

#Choose a model that fits on your machine

The model is a file several gigabytes in size: it must fit in your card's video memory, or in the RAM of a Mac or PC without a dedicated card. The Gemma 4 page in the Ollama library lists the variants and their download sizes. A variant that is too large for your machine will work, but slowly, spilling over onto the processor.

The right-hand column is our cautious estimate: add the context memory and a margin for the system to the sizes above. When in doubt, choose the smallest variant: a fast response from a small model teaches you more than a very slow response from a large one. You can switch models at any time.

#Your first session, step by step

Allow about fifteen minutes to go through the eight steps below. They cover everything from launching to verifying a response. Each step is independent: you can stop after the second and come back later. The only thing you lose by closing the session is the conversation history, which the terminal does not retain.

#1. Launch the model

ollama run gemma4:e2b

The ollama run gemma4 command appears in the Ollama documentation as an example of running a model. On the first launch, the model is downloaded: the duration depends on its size and your connection. Then a prompt appears, and the model waits for your message. Nothing happens until you type.

#2. Talk to it in complete sentences

First classic mistake: writing keywords as you would for a search engine. A language model does not search for documents; it continues text. So describe what you want, including the context and expected result.

# Trop vague
résumer réunion points importants

# Précis
Voici des notes de réunion. Résume-les en 5 puces : d'abord les décisions
prises, puis les actions à faire avec leur responsable.

Notes :
- Alice propose de migrer vers Postgres d'ici juin
- Bob rappelle qu'il faut finir l'audit sécurité avant
- Décision : on migre, mais l'audit passe avant (fin mai)
- Charles s'occupe de l'audit
- Alice s'occupe de la migration

A structure that works almost every time: the task, the data to process, and the expected format. Four exercises to practice: ask for a summary of personal notes, draft a three-line cancellation email, ask for a Python function that counts the words in a file, and ask a question the model can’t answer, such as the contents of a document you haven’t given it.

#3. Test what the model does not know

The final exercise is the most instructive. A model does not access the internet or your files unless a tool is connected: it knows only its training and the conversation text. Ask it for the contents of an imaginary document or about a recent event, and observe: a good model says it does not know; a less reliable model invents a plausible answer. This behavior is called a hallucination. The resulting rule: the more precise, recent, or specific the information is to you, the more you must provide it in the prompt or verify it. Our guide to hallucinations details the countermeasures.

#4. Fix it without starting over

Is the model responding beside the point? Don't just type “no, try again”: it has not understood anything more. Add a precise constraint that says what is wrong and what you expect.

Refais, mais en 3 puces maximum et sans jargon technique.
Chaque puce doit tenir sur une ligne.

Length

Number of chips, lines, or words.

Ton

Formal, casual, journalistic, instructional.

Form

List, table, JSON, email, code.

Prohibitions phrased positively

Prefer “start directly with the answer” to “don’t say here’s”.

#5. Session commands

In the session, lines that begin with a slash are not sent to the model: they control Ollama. The /? command displays the complete list.

/set parameter num_ctx 8192

Sets the context size for the session: Ollama’s FAQ cites this command to change the window.

/set system

Defines a system message, a permanent role for the session.

/show info

Displays information about the loaded model: architecture, size, context length, and quantization.

/clear

Clear the current conversation's context without unloading the model.

/bye

Close the session, or press Ctrl+D.

Once the session is closed, the model remains in memory for five minutes by default, according to the Ollama FAQ, which speeds up a new launch. To unload it immediately, run ollama stop followed by the model name; ollama ps lists what is loaded and shows the portion running on the GPU and CPU.

#6. Hold a multi-turn conversation

The model only remembers what is in the context window: on every turn, the application sends it the entire history. So you can chain requests that build on one another.

Je dois écrire un courriel pour annuler une réservation au restaurant.
Fais-le en 3 lignes, poli mais pas guindé.

[réponse du modèle]

Bien. Refais la même, mais plus chaleureuse, et précise que je
reprogrammerai dans 15 jours.

[réponse du modèle]

Parfait. Traduis-la en anglais.

The limitation is context size. According to Ollama's documentation, the default is around 4,000 tokens below 24 GiB of VRAM, 32k between 24 and 48 GiB, and 256k above that. When the history exceeds the window, the oldest exchanges are lost. If the model “forgets” the beginning of a long session, increase the window with the num_ctx command, keeping in mind that each context token consumes memory.

#7. Ask for code, and verify it

Code is a good testing ground because the result can be tested. Give it the language, input, expected output, and one constraint: “Write a Python function that takes the path to a text file and returns the number of words. Respond with the code only, followed by one explanatory sentence.” Copy the code into a file, run it on a case whose answer you know, and if an error appears, paste the complete error message into the conversation and ask it to fix it. This request, execution, correction cycle works better than demanding perfect code on the first try.

#8. Read a response critically

A fluent response isn’t an accurate response. The model produces plausible text; it doesn’t verify anything. Make a habit of checking anything that matters before using it.

Numbers and dates

Cross-check them against a source: these are the elements models most easily make up.

Names, references, and citations

Verify that they exist. A model can fabricate a book title or statute that sounds plausible.

Code

Run it before relying on it, and review what it does to your files.

Consistency

Ask the same question in two ways: if the answers differ, be wary of both.

A good practice is to provide the reference text yourself and ask the model to cite the passage that supports each claim. This lets you verify them in a few seconds instead of searching for everything yourself.

#Stay in the terminal or switch to an interface?

Start with the terminal to understand the principle, then switch to an interface as soon as you paste long texts or want to retrieve your conversations. The two work together: the model downloaded once with Ollama also serves the interfaces, with no second download, as long as they use the same local server.

#When something goes wrong

For finer control over creativity, the Modelfile documentation for Ollama summarizes the effect of temperature: the higher it is, the more creative the model; the lower it is, the more consistent it is. The guide to temperature and top-p details values to try.

#And after that first conversation?

A permanent role

The guide to system prompts shows how to give the model stable rules.

A more comfortable interface

Open WebUI or LM Studio add history, file attachments, and a readable layout.

Your own documents

RAG lets you query your files without pasting them in every time.

#Go further

• Install Ollama in 5 minutes
• Prompting basics
• Master system prompts
• Understanding the context window
• Hallucinations: how to limit them
• Open WebUI with Ollama: complete guide

• Source: Ollama FAQ
• Source: Ollama, context length
• Source: Ollama library, Gemma 4
• Source: Ollama, Modelfile reference