Lesson 4

How AI systems work

Connectors, memory, context, skills and tools

Wednesday 18 November · 14:20

Where we are

Lesson 4 of 12: How AI systems work

01 · Done

Introduction to Generative AI

Mon 9 Nov · 08:40, 10:20

02 · Done

AI for you: making it work in your day

Tue 10 Nov · 08:40, 10:20

03 · Done

Getting the most out of your AI tools

Wed 11 Nov · 08:40, 10:20

04 · You are here

How AI systems work

Wed 18 Nov · 14:20

05

AI at work and in teams

Wed 18 Nov · 16:20

06

Spotting AI opportunities

Fri 20 Nov · 08:40, 10:20

07

Practical business with AI

Mon 23 Nov · 08:40, 10:20

08

Ethical AI and what AI costs

Fri 27 Nov · 14:20, 16:20

09

Building with AI

Mon 30 Nov · 12:40, 14:20

10

Build and pitch prep

Fri 4 Dec · 08:40, 10:20, 12:40, 14:20

11

Final presentations

Tue 8 Dec · 08:40, 10:20

12

Reflection and final Q&A

Wed 9 Dec · 08:40, 10:20

Recap

In lesson 3, our main objective was to:

Get the most out of your AI tools: set them up, link them to your notes and ask them well

What we covered

Inside the LLM

It learns patterns and answers one token at a time.

Check its work

It predicts a likely answer, not a true one.

How AI remembers

A context window with a limit, and RAG to find your notes.

Better prompts

Who, what, where and when, and how.

Last week, in Lesson 3

Connect AI to your tools

The connector, and how far it can go

Last week

What is a connector?

A bridge that lets your AI read from, and act in, your other apps. Remember: where does the information live?

Answers use your real information, not guesses

It can act for you: add a task, draft an email

Only connect what you need, and check what it can see

Your AI app

Gmail

Calendar

Drive

Notion

Humble

Quick demo · With a connector

Claude reads my Humble notes

01 · YOUR ACCOUNT

My Humble

02 · ALREADY IN MY HUMBLE

Lecture 1: why AI, why now

· AI is spreading faster than any technology

· AI is not magic

· Five types of AI

Added yesterday

CONTEXT

Request for context

Search notes: “last lecture”

Connector

Checks your permission, then passes it on

Allowed

Sent back to Claude

Lecture 1: why AI, why now

“AI is spreading faster than any technology… Five types of AI.”

Claude

What did we cover in our last lecture? Quiz me on it.
Searched your Humble notes

From your note “Lecture 1: why AI, why now”: AI is spreading faster than any technology before it, and there are five types of AI. First question: which type of AI creates new text?

Reply to Claude…

What else can I do?

That’s cool, but what else can it do?

Try each one: copy the prompt into Claude, then look in your Humble.

Add a note

Add a note

Find your tasks

Find your tasks

Research and link

Research and link

Schedule a daily tidy-up

Add a note

01 / 04

Save the main points from this chat as a note in Humble, called “Lesson 2: AI for you”. File it in a folder called “ISM Essentials of AI”, and make the folder if it does not exist yet.

Humble makes the folder for you if it is new.

Find your tasks

02 / 04

Look at my recent chat conversations. Tell me the tasks I likely still have to do, as a list. Ask me to confirm, then create the ones I choose as tasks in Humble.

Your AI app needs to be able to search your past chats.

Research and link

03 / 04

Research Mahdi Shariff, co-founder of Humble AI, an AI tool that helps you organise your life as a personal second brain. He is also my teacher at ISM. Save what you find as a note in Humble and add him as a contact. Find his LinkedIn profile and link it to both, then link the note to the contact. Finally, find the page to install the Humble AI browser extension, and give me the link here in the chat, so I can see it in action when I open his LinkedIn.

Open his LinkedIn with the browser extension: your note and contact appear there.

Schedule a daily tidy-up

04 / 04

Set up a scheduled task that runs every morning at 8:00, and run it once now too. Each time it runs: 1. Review my recent AI conversations. 2. Make a Humble folder for each project or key area I am working on, and file my notes and tasks in the right one. 3. Keep a “Project brief” note in each folder up to date with the latest status. 4. Create notes and tasks for anything new, and tick off the tasks I have finished. 5. Send me a short summary of what changed and what to do today.

Find it afterwards under Scheduled in Claude.

Lesson 4 · New today

The 5 building blocks of AI

What every AI system is made of

Humble Explains

The 5 building blocks of AI

Think of AI like an intern. These are the five things it needs to work well for you.

Understanding the world

(The LLM)

01

What you ask it to do

(Prompt and instructions)

02

What it remembers

(Context)

03

Work it can repeat

(Skills)

04

Ways it can act

(Tools)

05

1 of 5

The LLM

Understanding the world

The brain. It learned how language and the world work from huge amounts of public text.

General knowledge: facts, ideas, how things work

It knows nothing about you or your work yet

It can be confidently wrong, so check it

Every other block builds on this one.

How it answers

“The capital of France is  ”

It guesses the most likely next word, one word at a time:

Paris

92%

Lyon

4%

the

2%

a city

1%

Each app runs on its maker’s own LLM. Numbers for illustration.

01 · The LLM

Where the LLM goes wrong

How it fails

It makes things up

It predicts a likely answer, not a true one.

It can trust a poor source

With search on, it can pick a wrong, old or weak page.

It can agree with you too much

It tends to go along with you, even when you are wrong.

Whose values are in it?

It is not neutral

People chose the data, rules and tuning. Those choices carry values, so an LLM is political.

Dominant languages win

Much of its text is in a few languages, mainly English. Those languages and their ideas get the strongest answers.

Some voices are missing

People offline, spoken cultures and small languages are thin in the data.

2 of 5

The prompt

What you ask it to do

The instructions for the job you want done right now: a question, a task or a request.

Clear and specific beats short and vague

Say who it is for and what you want back

It only lasts for this task

A prompt

Explain photosynthesis in three bullet points, for someone new to biology.
· Plants take in light, water and carbon dioxide.
· They turn them into sugar for energy.
· They give off oxygen as they do it.

One task, this chat only.

2 of 5

Custom instructions

A prompt for every chat

Instructions your AI follows in every chat, so you do not have to repeat yourself. Apps call this a system prompt.

Set it once, in your settings

Every new chat starts with it

Change it whenever your needs change

Next: where to find it.

Custom instructions

What should the AI know about you?

I am a first-year student at ISM, studying entrepreneurship. English is my second language.

How should it respond?

Short answers in plain English. Use bullet points. Ask me one question to check I understood.

Find it in Settings: look for Personalisation, Preferences or Saved info.

1

Click your avatar

You

Bottom left of ChatGPT

2

Open Personalization

Customize ChatGPT

Personalization

Memory

Plugins

Account

Email

mahdi@humbleai.com

Usage and limits

Putting it together

How does it actually work in practice?

An AI assistant

1You send a message

You

“Help me plan my study week.”

2Added for you

Your custom instructions

First-year student. Short answers in plain English.

3Everything is sent together

Custom instructions

“First-year student. Short answers in plain English.”

Your message

“Help me plan my study week.”

The AI model reads it all, every time

4You get a reply

ChatGPT

“Here is a simple plan: Monday, read. Tuesday, quiz yourself…”

Your words plus your instructions, every single time.

When it breaks

What happens when it all goes wrong?

An AI assistant

1You send a message

You

“Help me plan my study week.”

2Added for you

Your custom instructions

Law student. Long, detailed answers.

!

3Everything is sent together

Global instructions

“Short answers”

Local instructions

“Long, detailed answers”

Your message

“Help me plan my study week.”

Which instruction wins?

!

4You get a reply

ChatGPT

“A long law-style plan… even though you wanted short answers.”

!

Old or clashing instructions in, confused answer out.

As it grows

And complexity scales

Personal instructions

Set once in your settings

Project instructions

One set for each project

Team instructions

Shared with everyone

A shared team skill

Someone else’s version

A local skill

Your own copy

Your prompt

What you type right now

The AI model
reads all of it

Which one wins?

3 of 5

Context

Three places your AI gets context

THIS CHAT+ New chat
essay.docxGive me feedback
Your second point is strong…

Short-term memory

(The context window)

This chat and the files you attach. Gone when you start a new chat.

“Here is my essay, give me feedback”

Business course
notes.pdfbrief.docx
Memory: ISM student

Long-term memory

(Projects and saved memory)

What it keeps between chats: project files, instructions and things it remembers about you.

“Use my Business course files”

Connectors

(Your own tools)

Your email, calendar, files and notes, read when your question needs them.

“What is on my calendar this week?”

Long-term memory

How Retrieval Augmented Generation works

Prompt

A colleague asks a question
A user asks…
What are the spending limits and approvals needed based on the finance policy?

Retrieval

AI searches the knowledge base for parts that match
AI is searching…

Knowledge base

Finance policy
Strategy docs
Partner docs

Augmented generation

The LLM picks the text that fits the question
AI has found…

“Spending below £10k does not need approval.”

“Budget approvals from managers are needed for any spend.”

“Procurement approvals are needed for over £10k.”

“Procurement approvals can be made with this form.”

AI writes the answer from what it found
AI responds
The spending limit is £10k. You need approval from your manager, and from Josh in procurement using this request form.

3 of 5

Context

What it remembers

Short-term memory

(The context window)

  • This chat, and files you attach
  • It has a limit: long chats forget the start

Context

271k of 1M tokens

27% used

Your chat163k
Tools66k
Memory and skills28k
Instructions14k
Free space729k

“Here is my essay, give me feedback”

Long-term memory

(Projects and saved memory)

  • Files and instructions you keep
  • Ready in every chat in the project

Projects / Business course

Business course

How can I help you today?

Recents

Essay plan, week 3

Exam revision

Instructions

Be my study helper. Use UK English.

Memory

What it remembers from these chats

Files

Lecture 3 notes

PDF

Course guide

Drive

“Use my Business course files”

This is what we refer to as RAG (retrieval-augmented generation)

3 of 5

Context

Context from your own tools

Connectors let your AI read your email, calendar, files and notes. It looks only when your question needs them, and only with your permission.

Connectors

Gmail

On

Google Calendar

On

Google Drive

On

Humble notes

On
What is on my calendar this week, and are there any emails about it?

Looked in

Google CalendarGmail
You have 3 lectures and a team meeting on Thursday. Your teammate emailed yesterday asking to move the meeting to 14:00.

3 of 5

Context

You choose what it can do

For each connector, you set what your AI may do on its own and what it must ask you about first.

Always allow

It can use this without asking.

Needs approval

It asks you each time first.

Blocked

It can never use this.

Let it read. Make it ask before it sends or deletes.

← Connectors

Gmail

Tool permissions

⌄

Read-only tools

6Always allow

Get email message

Get email thread

Search email threads

⌄

Write and delete tools

24Needs approval

03 · Context

Where context goes wrong

The chat gets too long

The context window has a limit, so the start of a long chat falls out of view.

Old or clashing instructions

Two instructions disagree, so the answer is confused and nobody knows which one won.

The wrong note is used, or none

If it was never saved or connected, the AI cannot use it. An old or similar note can win.

It can see more than you meant

A connected tool can read private email, files and notes.

Humble explains

A lot of AI problems are
context engineering problems

Every AI system

InputsWhat goes in
+
TransformationHow the AI handles it
=
OutputsWhat comes out

Inputs

Impression

What the AI receives

You control this

InstructionsBackgroundExamplesLimits

Transformation

Guidance

How the AI handles it

You can shape this

System guidanceCustom instructionsOrderPriority

Outputs

Expression

What the AI produces

You see the result

ToneAccuracyDepthStructure

Context engineering

Workflows fail quietly when a step gets the wrong context

Trigger

New email arrives

Look up

Find the customer

Search

Search the help docs

1
AI step

AI writes a reply

2
Nothing found?
Stop

Flag it, do not guess

3
Check

A person approves

4
1

Look things up first, or the AI fills the gaps

2

Each step’s output becomes the next step’s context

3

Stop if nothing is found, or the AI will guess

4

A person checks before anything goes out

Context engineering

Every step and every handover can fail without anyone noticing

Trigger

New email arrives

Look up

Find the customer

Search

Search the help docs

AI step

AI writes a reply

Check

A person approves

It never triggers, so nothing starts

The info is missing, out of date or contradicts itself

The AI drops the key detail

It picks the wrong customer

The wrong info is passed on

The person is never told

The person makes the wrong call

4 of 5

Skills

Work it can repeat

Instructions packaged so the AI can reuse them, written in plain language in a text file.

Write the steps once

It uses them whenever the moment matches

Share them with your team

Skills turn a good prompt into a habit.

quiz-me.md

# Quiz me

Use when: I share lecture notes or slides.

Steps:
1. Pick the 5 most important ideas.
2. Ask me one question at a time.
3. Explain anything I get wrong.
4. End with a score out of 5.

Written once, used again and again.

04 · Skills

But what’s the challenge with skills?

Hard to share

Everyone has a different version

A prompt library grows fast

A local skill clashes with a global one

5 of 5

Tools

Ways it can act

Abilities the AI can use to do things, not just say things: search, code, make files and work in your apps.

Each tool does one kind of job

Connectors bring in tools for your own apps

More tools means more care with permissions

Brain, prompt, context, skills and tools: that is an agent.

Things it can use

Search the web

Run code

Make files

Your calendar

Your email

Your notes

You decide which tools it may use, through connectors and permissions.

05 · Tools

The challenge with tools

It can act, not just say

A tool can send, delete or buy, so a mistake costs more than a wrong sentence.

More tools, more to manage

Each tool needs access. The more it can reach, the more can go wrong.

It can pick the wrong tool or source

It can use a wrong or old result, and still sound sure.

Every step uses tokens

An agent that uses many tools can cost far more than a short chat.

Big questions

When AI acts for us

There are no right answers yet. Pick one with your table and say what you think.

1

Who decides what an AI tool is allowed to do in a company?

2

If an AI tool makes a mistake, who is responsible?

3

If a tool talks like a person, how do we know it is not one?

4

If a tool works for someone, how do we know who?

5

Should an AI tool get its own name and account, like a new team member?

Recap

The 5 building blocks of AI

The LLM

The brain: what it knows and how it reasons.

“Claude, GPT or Gemini models”

Prompt or instructions

What you ask the agent to do.

“Make a slide deck on renewable energy”

Context

Information it can see and use.

“Here is an example deck I want you to copy”

Skills

Instructions packaged so it can reuse them, written in plain language in a text file.

“A slide-deck skill it uses whenever you ask for slides”

Tools

Things it can use to act: apps, search and code.

“Web search, your calendar, a spreadsheet”

Recap

What we covered in Lesson 4

Connect your tools

The connector lets AI use your notes and apps.

5 building blocks

LLM, prompt, context, skills and tools.

Memory

Short-term is this chat. Long-term is found with RAG.

Context engineering

Inputs + transformation = outputs.

Workflows

Each step’s output is the next step’s context.

Big questions

Who is responsible when AI acts for us?

Hands-on · 5 minutes

Save your takeaways

Use your AI tool and the Humble connector to keep what you learned today.

Save a note in Humble called “Lesson 4: How AI systems work”, with my three biggest takeaways from today: [type them here]. File it in a folder called “ISM Essentials of AI”, and make the folder if it does not exist yet.

Lesson 4: How AI systems work

New note · ISM Essentials of AI

Up next

Lesson 5: AI at work and in teams

How AI changes your job and what stays human, then what changes when a team uses AI.

Wednesday 18 November, 16:20

Open the Lesson 5 slides

← Lesson

1 of 34