IntroductionBeginner thirty minutes · hands on

Tacking UpAI in thirty minutes

We are going to plan your Saturday. Along the way you will see what an AI model is on its own, what changes when it gets tools, and what people mean when they say “agent”.

By the end of thirty minutes you will have a real plan for this Saturday in your own town, you will be able to explain to someone else why the first answer was useless and the last one was not, and you will have asked it one real question about your business.

Presented by Emile du Toit
brainitconsulting.com

Before we start — two minutes

Open claude.ai on a laptop or phone and sign in. No account? Click Sign up and use your email. You will also need your mobile phone: Claude texts a code to it, and there is no way round that step. It is free and needs no card.
Free accounts get a limited number of messages every five hours, and each search uses some. If Claude says you have hit your limit, watch the screen with me and try the last two prompts at home.
Type your town and your line of work below. Every prompt on this page fills them in for you. Then each prompt has a Copy button: click it, paste into Claude, press Enter.

Your town What your business does

This page keeps it in your browser. It only goes to Claude when you paste a prompt.

The horse in the field

Five minutes. What the model is, with nothing on it.

Picture a horse in a field. It is strong, it is quick, it has covered a great deal of ground in its life. And right now it is doing nothing for you, because it has no bridle, no saddle, and no reason to go anywhere.

That horse is a large language model, an LLM. Claude, ChatGPT and Gemini each run on one. All it does is read the words in front of it and work out what word comes next, again and again, until the answer is finished. It learned how by reading an enormous amount of writing. Let us see what that gets you.

First, take the tack off

Start a new chat. Before you type anything, click the + at the bottom left of the message box. Click Web search and then Memory, so neither has a tick beside it. Memory lets it bring in what it knows about you from earlier chats, and that is tack too. Now paste this:

What is on in [your town] this Saturday?

What you will see

It has nothing to look with. When we tried it (September 2026), it gave no list at all, just a note saying Web search is off and a button to turn it back on. A model with no tools can also guess instead, and hand you a confident list from a year it half remembers. Either way, it cannot know.

Fig. 1 — The model on its own. Text goes in, text comes out, and that is its whole connection to the world.

When it makes something up and sounds sure of itself, that is called a hallucination. The horse shied at a shed snakeskin: confident, and wrong about something harmless. Here it was harmless. On your tax figures it would not be.

A horse in a field is not transport. It becomes transport when you tack it up.

Putting the tack on

Seven minutes. Same horse, with something to carry.

Nobody went and fetched a cleverer horse. They put a bridle on this one so they could steer it, a saddle so every ride starts the same way, and saddlebags so it can carry things back. The industry word for all of that together is the harness: everything put around the model. Never one strap. All the tack at once.

Fig. 2 — Same horse, twice. Everything that changed is ember, and none of it is the horse.

The piece that matters most is the saddlebags, the tools. A tool is a named thing the model is allowed to ask for: search the web, read a file, look at a calendar. The model only ever asks. Software you control does the doing, or stops and checks with you first.

Fig. 3 — Tool use. The model never reaches into anything itself. It asks, and something you control decides.

Now put the saddlebags on

Start a new chat, so the first answer does not colour this one. Click + and switch Web search back on. Then ask the same thing, plus one line:

What is on in [your town] this Saturday, and what will the weather be like? Give me a link for each thing you tell me.

Same question, same model, a completely different answer, because this time it could go and look. Click one of the links. Always ask for them. An answer with a source you can check is a different thing from one you have to trust.

Ask it which bag it reached into

Which tools did you just use to answer that, and what exactly did you search for?

It will name them, and the names are dull. When we tried it (September 2026), it listed three: web_search for the events, weather_fetch for the forecast, and web_search_fast to find a forecast page it could link to. It also said it had ignored search results showing an old winter forecast, and that show times were worth checking on the linked pages. Nothing mysterious happened. It was given reach, and it told you how it used it.

The ride

Seven minutes. One question is a trick. A job is a loop.

So far you asked, it answered, done. That is a horse doing one trick. Give it a job with several parts and something different happens: it looks, it decides what to do next, it does it, it checks what came back, and it goes round again until the job is finished.

That is the agentic loop, and it is all that people mean by an AI agent. Not a robot. A model with a harness on, a goal from you, and permission to keep going round.

Fig. 4 — Round and round until the goal is met or it needs you. You set both.

Give it the whole job

Web search still on. Paste this, then watch the screen while it works rather than waiting for the end. Claude shows each search as it happens. When we tried it (September 2026), the weather came first, then about six searches.

Plan my Saturday in [your town]. Check the forecast first. Then find one thing to do in the morning, somewhere for lunch that is open on Saturdays, and one thing for the afternoon. If rain is likely, keep the afternoon indoors. Give me a timed plan with a link for each stop.

Count the laps

Look. Did it check the weather before anything else? You told it to.
Decide, act. Count the separate searches. Each one is a trip round the loop.
Check. Did it change anything because of what it found — rain, a closed restaurant?

Now make it check its own work

Go back over that plan. Is every place actually open at the time you put it in? Fix anything that is not, and tell me what you changed.

When we tried it (September 2026), the first plan had the River Arts District until 5 PM, and it had never checked studio hours. Asked to check, it looked them up, found some studios close at 4, and moved the visit earlier. That is the check step made visible, and it is the most useful habit it has.

In a chat, you are still in charge of every lap: it cannot do anything except search and write back to you. Before it could book the table or text your friends, it would stop and ask. That stop is the gate, and it matters most in the next chapter.

The loop is the agent. The gate is what keeps it yours.

Words back, or work done

Five minutes. Watch this one on the screen.

Everything so far gave you words back. Claude can also take a job and a folder and hand back finished files. Some paid accounts show this as a separate Cowork button in the message box. On newer paid accounts there is no button: you just ask for files in the same chat. Either way it needs a paid plan, so I will drive this part and you watch. The prompt is here if you want to try it later.

On the screen: the same Saturday, done in Cowork

An empty folder called saturday, connected in Cowork, set to Manual, so it asks before it saves anything. One prompt:

Plan my Saturday in [your town], checking the forecast first. Save three files into this folder: plan.txt with a timed plan and a link for each stop, bring.txt with what to take given the weather, and invite.txt with a short, friendly text inviting two friends along. Ask me before you save each one.

Watch for three things.

The same loop as chapter three: forecast first, then the searches.
The gate. Before each file is saved it stops and asks. Nothing lands in the folder until I say yes.
Real files. Open the folder afterwards. They are on the computer, and they will be there tomorrow.

It wrote the invitation. It did not send it. Sending would be another gate, and that one stays with you.

When we tried it (September 2026)

It checked the forecast first, with the National Weather Service: sunny, a high near 71°F. Then it stopped and asked before saving each of the three files, one at a time. What landed in the folder:

plan.txt: breakfast at 8, a short hike on the Blue Ridge Parkway, lunch downtown, the River Arts District, tapas at 7:30, a link for every stop, and a reminder to check Parkway road closures before leaving.
bring.txt: layers for a cool morning up on the ridge, sunscreen, and an offline map, because phone signal is patchy on the Parkway. Nobody asked for that. It worked it out from the weather and the route.
invite.txt: a few friendly lines, ready to paste into a text.

One detail worth noticing: in Chat, the weather tool had said 74°F for the same day. Two sources, two slightly different numbers. That is harmless on a Saturday and exactly why you check the ones that matter.

	Chat	Cowork
You hand it	A question, or something you paste or attach	A job, and a folder or connected account to do it in
You get back	Words, which you copy somewhere	Finished files, and a summary of what it did
Before anything real	It can only answer you	It stops at the gate and asks
Plan	Free	Paid plan. Folders on your computer need the desktop app

Words you will paste somewhere: Chat. Finished work: Cowork.

There is a third door, Code, for building software. It is a good door and not today’s. That is what the twelve-session series is for.

Now ask it about your business

Three minutes. Same loop, real work.

Everything you just did to plan a Saturday works on a Monday. The loop does not care whether it is looking up a lunch spot or a competitor’s prices. Web search still on, one more paste:

I run [what your business does] in [your town]. Find five businesses near me that do the same thing. For each, tell me what they charge if it is published, what their customers praise, and what they complain about. Then tell me one thing I could do differently. Link every fact.

Read it the way you would read a report from a new employee: click two of the links, check one price, and notice anything it could not find. It is good at the digging. You are the one who knows whether the conclusion is right.

What to keep away from it at first

Anything you cannot check. If you could not spot a wrong answer, do not hand the job over yet.
Sending on your behalf. Let it draft. You send.
Money, and other people’s private details you would not email to a stranger.

Questions people actually ask

The ones that come up in the first thirty minutes, every time.

Is this the same as ChatGPT?

Different stable, same kind of horse.

Same idea underneath: a model, a harness, tools and a loop. The buttons have different names. Get comfortable with one before you compare them.

Will it send things or book things without asking?

Only through a gate you opened.

In Cowork it stops and asks before it acts, as long as you leave the approval setting on Manual, which is where it starts. Leave it there until you know a job well.

Does it remember me next time?

A fresh horse has never met you.

A new chat starts from nothing except what the app gives it. If you keep repeating the same background, put it in a Project with standing instructions, and every chat in that project starts with it.

It got something wrong. Now what?

It shied at a snakeskin.

It will, sometimes, in the same confident voice as when it is right. Ask for links, check what matters, and when it is wrong, tell it what is wrong in one sentence. That is the normal way of working.

Is what I type used to train it?

Decide once, in the settings.

It is a setting in your account. Open Settings, then Privacy, and set Help improve our AI models the way you want it.

Plain words

Every term from today, each with its horse.

Large language model (LLM): The engine that predicts the next word. The horse.
Hallucination: Making something up and sounding certain. Shying at a shed snakeskin.
Harness: Everything put around the model: the app, instructions, tools and approvals. All the tack at once.
Tool: A named action the model may ask for, which software you control carries out. A saddlebag.
Agent: A model given a goal, tools, and permission to keep going. A working horse.
Agentic loop: Look, decide, act, check, repeat. The ride itself.
Permission prompt: The stop before it changes or sends anything. The gate, and your hand on the latch.
Chat and Cowork: Two ways of working in the Claude app: one gives you words back, one takes a job and a folder and hands back finished files. The field, and the working horse in your room.

Everything on one page

Photograph this before you go.

1 · The horse

It predicts the next word. On its own it cannot look anything up or do anything.

2 · The tack

The harness. Tools are the difference: it asks, the app does.

3 · The ride

Look, decide, act, check, repeat, and stop at the gate before anything real.

4 · Words back, or work done

Chat when you want words back. Cowork when you want the work done.

Before you leave

Pick one job from last week that was mostly looking things up, reading or writing, and that annoyed you. Tonight, give it to Chat the way you gave it your Saturday: the goal, what to check first, and ask for links.

If you want more

#001 — The Horse in the Field

Today was the horse, the tack and the ride in thirty minutes. The series takes each one slowly, with your own business on the page: steering it properly, writing standing instructions once, and then, a session at a time, giving it real work.

Before you start it: nothing. The free account you used today is all it needs.

Want someone to hold the reins with you?

I do this for a living, and love to help people.

Thirty-plus years in enterprise software and a computer science degree — which mostly means I have watched a great many horses bolt, and I can usually tell you which ones are worth saddling before you spend money on the saddle.

In person

Your team, your room, your actual work on the whiteboard instead of somebody else's examples.

Over video

The same session, at your desks, with fewer chairs to stack afterwards.

Built for real

When the idea survives the workshop and somebody has to go and build the thing — that part I do too.

Fun fact — I rode to school on horseback as a kid in South Africa. So the metaphor isn't borrowed. I've done the falling off in person.

Emile du Toit brainitconsulting.com

Introduction · Tacking Up