Skip to content
Coffee?

My workshop

What I am learning, connecting and testing.

Smart Flows · concepts

Anatomy of AI

I am learning these concepts and arranging them to understand how they relate.

I am not trying to explain AI as an expert. This is my way of understanding it: when a new word comes up, I look at what it does, what it resembles and what it connects to.

I arrange it as a journey, from what we ask to what we get. And alongside it, always present, human judgement.

Human judgement

How we talk to it

Decides what to ask and what context to give

Who answers

Modelthe engine that interprets and generates

How it acts

Sets the limits

How it connects and looks things up

Chooses what it can access

What we get

Resultwhat we receive

Reviews, corrects or decides

Tap a piece to open its card.

Provisional illustration

How we talk to it

Prompt

The instruction and context we give a model to guide its response.

Put simply

It is what we ask the AI and how we ask it. The clearer the request, the more likely the answer is to match what we need.

The image

The note of instructions you leave on the desk before work begins.

What it solves

It steers the answer towards a specific goal, tone and format.

What to watch out for

A good prompt does not guarantee a good answer. What comes back still needs checking.

Related to

I test it in Prompt Control →
Provisional illustration

How we talk to it

Context

The information the model receives to better understand what it should do.

Put simply

AI knows nothing about our situation except what we tell it or give it to read: documents, examples, earlier conversations.

The image

The work table: only what is on it can be used.

What it solves

Answers that fit our reality better and are less generic.

What to watch out for

Too much context also confuses, and it should never include confidential data without permission.

Related to
Provisional illustration

How it acts

Agent

An AI system that can chain steps, use tools and act to reach a goal.

Put simply

Instead of answering a question, it receives a goal and decides which steps to take to reach it.

The image

An assistant with a list of errands who decides in which order to run them.

What it solves

Multi-step tasks that used to require asking for one thing at a time.

What to watch out for

The more autonomy it has, the more important it is to know what it does at each step and to be able to stop it.

Related to

I test it in AI vs AI →
Provisional illustration

How it acts

Workflow

An organised sequence of tasks or decisions that make up a process.

Put simply

A route set in advance: first this, then that. AI can take care of some steps, but the process sets the order.

The image

A recipe: the same steps, in the same order, every time.

What it solves

Repetitive tasks that gain speed and consistency.

What to watch out for

An error in one step carries into the next ones. Checkpoints are worth planning.

Related to

I test it in Research Flow →
Provisional illustration

Human judgement

Human in the loop

An automated process in which a person keeps checkpoints for review, decision or validation.

Put simply

The machine moves forward, but at certain points it stops and waits for a person to look, correct or approve.

The image

The signature before sending: nothing goes out without someone looking at it.

What it solves

It stops errors from travelling far and keeps responsibility in human hands.

What to watch out for

If everything is reviewed, the process slows down. If nothing is, it loses its point. The key is choosing the right moment.

Related to

I test it in Human in the loop →
Provisional illustration

How it connects and looks things up

API

Application Programming Interface. A standardised way for two applications to exchange information or carry out actions with each other.

Put simply

It is the channel through which one tool asks another for something: looking up data, creating a document, sending a message.

The image

The counter window of an office: you ask using a specific form and get an answer.

What it solves

Different tools working together without anyone copying and pasting between them.

What to watch out for

Every connection opens a door. You need to know which permissions are granted, and to which data.

Related to

I test it in Behind the scenes →
Provisional illustration

How it connects and looks things up

MCP

Model Context Protocol. A protocol for connecting AI models to tools, files and other sources of information in a structured way.

Put simply

A common language that lets AI use many different tools without needing a special cable for each one.

The image

The universal travel adapter: one format to plug in many things.

What it solves

Connecting AI to documents, folders or apps in an orderly, reusable way.

What to watch out for

Easy to connect does not mean everything should be connected. Better to give access only to what is needed.

Related to

I test it in Behind the scenes →
Provisional illustration

What can go wrong

Hallucination

An AI response that sounds plausible but contains incorrect or made-up information.

Put simply

AI does not “know” that it does not know: when information is missing, it sometimes fills the gap with something that sounds right.

The image

The very confident storyteller who fills the gaps in their memory with imagination.

Why it matters

Understanding it helps us not to confuse confidence with accuracy.

What to watch out for

Data, quotes, figures and sources: anything that looks like a fact is worth checking.

Related to

I test it in AI Quality Check →
Provisional illustration

Human judgement

Guardrails

Rules and limits that define what an AI can do or answer.

Put simply

They mark out the playing field: topics that are off-limits, actions that need permission, formats to respect.

The image

The railings along a mountain path: you move freely, but you stay on the track.

What it solves

It reduces risks without having to supervise every step.

What to watch out for

No barrier is perfect. Limits are reviewed and adjusted through use.

Related to

I test it in AI Quality Check →
Provisional illustration

How it connects and looks things up

RAG

Retrieval-Augmented Generation. A technique that lets an AI consult external information before generating a response.

Put simply

Before answering, the AI searches a specific collection of documents and relies on what it finds.

The image

Checking the library before answering.

What it solves

Answers based on your own, more up-to-date sources, with less invention.

What to watch out for

If the documents are out of date or poorly chosen, so will the answer be.

Related to

Quick glossary

Definitions, in one line

API

Application Programming Interface. A standardised way for two applications to exchange information or carry out actions with each other.

MCP

Model Context Protocol. A protocol for connecting AI models to tools, files and other sources of information in a structured way.

Agent

An AI system that can chain steps, use tools and act to reach a goal.

Workflow

An organised sequence of tasks or decisions that make up a process.

Connector

The bridge that lets one tool access another application or source of information.

Human in the loop

An automated process in which a person keeps checkpoints for review, decision or validation.

Prompt

The instruction and context we give a model to guide its response.

Model

The AI system that interprets an input and generates a response.

Context

The information the model receives to better understand what it should do.

Hallucination

An AI response that sounds plausible but contains incorrect or made-up information.

RAG

Retrieval-Augmented Generation. A technique that lets an AI consult external information before generating a response.

Token

The unit into which a model splits and processes language.

Guardrails

Rules and limits that define what an AI can do or answer.

← Back to Smart Flows

Audiovisual content created with AI from my real image.