Skip to content

My Virtual Atelier

Virtual café

AnatomIA

An interactive way to explore AI concepts and reflect on how we use them.

AnatomIA

My map for understanding AI

A person interprets through lived experience. A language model relates data to learned patterns to generate a response.

Choose a tab to explore each section.

What does AI see?

What do you see?

Before continuing, take a few seconds to look at the scene.

A woman leaves a house with a pale façade, a wooden door, a window and plants, in warm light.

The labels are preset examples. This experiment does not analyse the image in real time.

What does SofIA advise us?

A reminder to use AI while keeping your own judgment.

SofIA

SofIA · created with AI, guided by human judgment.

Two ways to answer a question

AspectPersonDiagram of a human brain with connections and a close-up of neurons and synapses.AILanguage modelArtificial network with nodes, layers and a flow of input and output data.
InputSenses, body and environment.Data and task instructions.
What informs the responseMemory, emotions and lived context.Patterns learned during training and available context.
How it respondsInterprets through experience.Generates a response using patterns and probabilities.
Who decidesAssesses the response and bears the consequences.Its output requires human assessment.

Simplified comparison. An artificial network does not reproduce a person’s lived experience.

From question to response

  1. What we need to obtain.

  2. How we frame the request.

  3. The information we make available.

  4. Which instructions we reuse for this task.

  5. The system that processes the information.

  6. The actions it can take.

  7. How it accesses other resources.

  8. The output it generates and what can go wrong.

  9. What we check before using it.

The route depends on the task. Tools and connections are optional and can provide context before or during generation.

Where the person still matters

Judgment cannot be automated.

Intention
Define why we are asking the question.
Context
Supply what is missing and put the response in perspective.
Interpretation
Make sense of the response through experience and context.
Review
Check data, sources and assumptions.
Decision
Choose what to use and what to discard.
Responsibility
Be accountable for how we apply the result.

Which model should I choose for a task?

Explore Model Choice →

What is a token?

A unit a model uses to process text. It may be a word, part of a word or a punctuation mark.

Explore further

When an action is triggered

Hooks

A configured action triggered by an event in the workflow, such as before or after using a tool.

What it is for

Link a check or a notification to a specific point in the process.

What to review

Its behaviour depends on the environment and configuration. Understand what it runs and review the result.

Example

After editing a file, run a formatting check.

Hooks · Claude Code ↗

How it connects and consults

Connector

A bridge that provides access to an application or data source.

What it does

Connect a document folder to consult files.

What to watch

Check permissions: what can be read, changed or sent.

Examples

A folder of files accessible through a connector.

How we talk to it

Token

A unit a model uses to process text. It may be a word, part of a word or a punctuation mark.

What it does

Measure the amount of text the model processes or generates.

What to watch

One word does not always equal one token. The split depends on the model and language.

Examples

A long word can take several tokens. Punctuation is processed too.

OpenAI documentation ↗

How we talk to it

Prompt

The instruction given to a model to guide its response.

What it is for

Define the task, objective, tone or format.

What to watch

A good prompt does not guarantee a good response. Context and verification matter too.

I test it in Prompt Control →

Example

“Summarise this report in five points.”

How we talk to it

Context

The information AI receives to understand the situation.

What it is for

Give the model what it needs to produce a more relevant response.

What to watch

Incomplete, excessive or incorrect context can steer the response in the wrong direction.

Example

An email from a customer and the terms of her order.

How we instruct it

Skill

A reusable set of specialised instructions.

What it is for

Tell AI how to carry out a type of task consistently.

What to watch

A skill guides behaviour, but does not replace the model or human judgement.

Example

A reusable guide for preparing summaries with a fixed structure.

Who responds

AI

The general term for systems capable of performing some tasks associated with human intelligence.

What it is for

Analyse, predict, generate, classify or automate, depending on the system.

What to watch

AI systems do not all work in the same way or have the same capabilities.

Example

A system that identifies objects in photographs.

Who responds

LLM

An LLM is a type of model focused on understanding and generating language. LLM describes a category, not a brand.

Examples

The GPT family provides examples of language models. ChatGPT is the application through which we interact with models.

OpenAI documentation ↗

What it is for

Understand and generate language, summarise, rephrase, analyse or converse.

What to watch

It produces a plausible response based on how it works, which is not necessarily the truth.

Who responds

Model

A system trained to produce results from data. An LLM is a type of model.

Examples

GPT-6.1 Sol is a specific model available in ChatGPT Work and Codex. A model that classifies images is also a model, but not necessarily an LLM.

OpenAI documentation ↗

What it is for

Interpret the information received and produce a prediction or response.

What to watch

Models differ in their capabilities, accuracy and limitations.

Who responds

Local AI

AI that runs on your own computer or server.

What it is for

Process data without necessarily sending it to an external service.

What to watch

Privacy does not mean simplicity. Hardware, maintenance and security still need to be managed.

Example

A model installed and run on a laptop.

How it acts

Agent

A system that can perform a sequence of actions to reach an objective.

What it is for

Plan, use tools, consult information and carry out several steps.

What to watch

The more autonomy it has, the more important limits and controls become.

I explore it in Human in the Loop →

Example

Find a file, prepare a draft and wait for your validation.

How it acts

Workflow

A sequence of steps to complete a task. It can be manual, partially automated or automatic.

What it is for

Structure a process and automate some repetitive actions.

What to watch

Automating a poor process does not improve it.

Example

Receive a form, classify the request and notify a person.

How it acts

Tool

A function AI can use to act beyond the conversation.

What it is for

Search, calculate, read a file, send information or trigger an action.

What to watch

Control what the tool can do and which data it can access.

Example

A calculator that adds several figures.

How it connects and retrieves

API

An interface that allows two software applications to communicate.

What it is for

Exchange data or trigger functions between different applications.

What to watch

Permissions, the data exchanged and secure access.

Example

One application asks another for an order’s status.

How it connects and retrieves

MCP

A protocol that standardises connections between AI, tools and data sources.

What it is for

Connect different systems without rebuilding every integration from scratch.

What to watch

Keep the connection under control. Who can access what, when and with which permissions.

Example

A server offers tools for consulting files through MCP.

How it connects and retrieves

RAG

A method that retrieves information before generating a response.

What it is for

Make AI respond using specific documents or knowledge.

What to watch

If the retrieved information is wrong or incomplete, the response may be too.

I explore it in Research Flow →

Example

Find passages in a document to support an answer.

What we get

Result

What the system produces at the end of the process.

What it is for

Text, analysis, recommendation, file, intermediate decision or action.

What to watch

A well-written result is not necessarily correct.

Example

A summary of a report’s main points.

What we get

Hallucination

When AI generates plausible information that is false or invented.

What it is for

It is not a function. It is a risk to understand.

What to watch

The more important the information, the more necessary it is to verify its sources.

I explore it in AI Quality Check →

Example

An invented reference that looks authentic.

Human judgement

Guardrails

The rules and controls that define what AI can do.

What it is for

Limit actions, access or certain system behaviours.

What to watch

They reduce risks, but never eliminate them completely.

Example

Ask for approval before sending an email.

Human judgement

Human in the loop

A person intervenes at a key point in the process.

What it is for

Verify, correct, arbitrate or authorise a decision before it takes effect.

What to watch

The person should intervene where their judgement adds value.

I explore it in Human in the Loop →

Example

Review a draft and approve it before it is sent.

This website includes audiovisual content created with AI.