The Horse
in the Fieldwhat a chatbot actually is
Artificial intelligence explained through something people have understood for four thousand years: a horse, a saddle, and a job that needs doing.
By the end of this one you will have an account open, you will know how to steer it with a firm rein instead of a loose one, and you will have written your standing instructions down once so you stop repeating yourself every morning. Four exercises, and you can do all of them on your own.
Presented by Emile du Toit
brainitconsulting.com
Contents
- Before you startWhat you need, and opening an account. Five minutes.
- The horse in the fieldWhat it really does, and what it cannot do.
- Bridle and reinsThe chat box. Steering with words.
- Saddle and stirrupsStanding instructions that make a ride repeatable.
- When the horse boltsHow it goes wrong, and how to sit it.
- Where to startTwo websites, and the one setting worth finding.
Then the questions everyone asks, a plain-words glossary, and a one-page picture worth photographing before you leave.
Five short exercises are folded into the chapters. Open one and it tells you exactly what to type.
Before you start
- Any computer with a web browser. Windows, Mac, even a tablet for this one. Nothing gets installed today.
- An email address you can open. You will be sent a link or a code to sign in.
- About 45 minutes, and somewhere you can think for ten of them.
- Facts about your own business, if you have one — what you do, what you charge, what you refuse. You will use them in the third exercise. If you do not have a business, borrow Dale's; his numbers are in chapter three.
No card, nothing to buy, nothing to install. If you have never done anything like this, you are exactly who this was written for.
Lab 0 — Open an account 5 min
Everything after this needs an account, so do it now and read the rest with the window open beside you. Both are free. Pick one — you do not need both, and comparing them today will only slow you down.
Either one of these
- Go to chatgpt.com or claude.ai.
- Click Sign up. Use your email address, or the “continue with Google” button if you would rather not think of another password.
- Check your email. Click the link, or type in the code if it sends you one.
- You will land on a page with a big empty box at the bottom. That is the whole thing. That box is the reins from chapter two.
Say hello, so the first message is out of the way
Type this and press enter. It does not matter what it says back — the point is that the box works and you have now used it once.
If it asks you to choose a plan, look for the word Free. You never need to enter a card for anything in this workshop.
One setting, while you are here
Open Settings and find the option about whether your conversations may be used to improve the model. Both services have one, both make it easy enough to find, and both default differently. Decide once now rather than wondering later. Chapter five comes back to this.
The horse in the field
Powerful, fast, well-traveled — and standing there 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 absolutely nothing for you, because it has no bridle, no saddle, and no reason to go anywhere.
That horse is what the industry calls a large language model — an LLM. ChatGPT, Claude and Gemini are all built on one. Before anyone puts tack on it, here is what it actually does:
It reads the words in front of it and works out what should come next. Then it does that again, and again, one word at a time, until the answer is finished. It learned how by reading an enormous amount of writing — books, manuals, websites, conversations.
To guess the next word well enough to write a decent estimate letter, you have to have absorbed how estimates go, what a customer worries about, and what ordinary business writing sounds like. The trick is simple. What the trick requires is not.
Four things to hold on to
- It is good at language work. Drafting, summarizing, rewriting, translating, explaining, sorting, pulling the useful line out of a long email.
- It knows nothing about your business. Not your prices, not your customers, not your rules — unless you tell it, every time.
- It has no memory of yesterday. A fresh chat is a horse that has never met you.
- It cannot do anything. It produces text. It cannot send the email, open the spreadsheet or take the payment. Not yet — that’s next session.
A horse in a field is not transport. It becomes transport when you put a saddle on it.
Bridle and reins
Add the reins and you have a chatbot. Now it goes where you point it.
The bridle is the chat box. You type, it answers — that back-and-forth is what people mean by a chatbot. Model plus reins. Nothing more mysterious than that.
Reins only work if you use them. A loose rein gets you a horse wandering toward whatever looks interesting, which is exactly what a vague question gets you: a general, hedge-everything answer that could have been written for anybody.
Loose rein, firm rein
You get four paragraphs of generic advice.
Write something about our new opening hours.
You get something you can send.
Write a 60-word notice for our front window.
We're open Sundays 10–4 starting March 1.
Friendly, no exclamation marks.
We're a family-run bakery in Asheville.
Four things did the work there: who you are, what you want, how long, and what tone. That is most of what people dress up as “prompt engineering”.
Lab 1 — Loose rein, firm rein 10 min
Do not take my word for the difference. Ask the same thing twice, badly and then properly, and read the two answers side by side. This is the exercise people tell me about months later.
First, the loose rein
Type this exactly. It is deliberately vague.
Read what comes back. It will be competent, polite, and about nobody. Notice how much of it you would have to delete.
Now the firm rein
Start a new chat, so the first attempt is not colouring the second. Then write the same request with the four things that were missing — who you are, what you want, how long, and what tone.
Then do it again with your own business. Your trade, your real change, your real town. That version is the one worth keeping.
What to notice
- The second answer is usable and the first is not, and the only difference is what you told it.
- It never asked you for the missing information. It guessed, confidently. That is chapter four arriving early.
- You did not learn a trick. You wrote a clearer brief — the same thing you would have to do for a new employee.
If the second answer is still not right, do not start over. Tell it what is wrong in a sentence — “too formal, and cut the last line” — and let it try again. That back-and-forth is the normal way to work, not a sign you asked badly.
How far it can see: the context window
Everything in the current conversation is what the horse can hold in its head on this one ride — the context window. It is large but not endless. On a very long ride, the earliest part of the conversation falls out the back, and the model stops referring to it.
Two practical habits follow. Start a new chat when you change subject. And when something matters, don't rely on it having been said an hour ago — say it again.
Saddle and stirrups
Bareback is fine for five minutes. A day's work needs a fitted saddle.
Anyone can sit on a horse for five minutes. Do it for a day and you will want a saddle, stirrups, and a girth that stays tight. In AI, the saddle is the set of standing instructions that goes along with every message, whether you type it or not.
The technical name is the system prompt. You will meet it under friendlier names: a Custom GPT (ChatGPT), a Project (Claude and ChatGPT), a Gem (Gemini), or plain “instructions” (all three). Same saddle. You fit it once and every ride after that starts properly.
What goes in a saddle
Who you are. You help Dale & Son, a two-truck residential plumbing company, turn inquiry emails into written estimates.
What you always do. Pull out the job, the address and the access notes. Price it from the facts below. Show labor and parts on separate lines, with the service call on its own line.
Facts you may rely on. These are the only numbers you may use.
- Service call — $95, covers travel and the first hour on site.
- Labor after the first hour — $110 an hour, in half-hour steps.
- Saturdays and evenings — the labor rate ×1.5. The service call does not change.
- Parts — what we paid, plus 20%.
- Sales tax — 7%, on parts only. Never on labor.
- We do not work on gas lines.
Tone. Plain, warm, no jargon. Short sentences.
Never. Never quote for gas work — flag it for Dale. Never promise a date. Never guess at a price you were not given: say what you'd need to know.
Three inquiry emails to try it on
Paste one of these in after the instructions are set. The third one is there on purpose.
What to look for in the three answers.
- Marisol is the straightforward one. Service call, labor, the heater as a part. Check the arithmetic yourself, once.
- Ken should come back at the Saturday rate on the labor and the ordinary rate on the service call. If it applied ×1.5 to the whole thing, your instructions were not specific enough — and now you know exactly which line to tighten.
- Dee is the test. It must refuse to price the gas line and flag it for Dale, while still quoting the slow drain. If it prices the gas work, the “never” did not take, and that is worth finding out on a made-up email rather than a real one.
None of them says how old the pipework is or whether there is a shutoff, so a good answer will also tell you what it would need to know. That is the last line of the instructions doing its job.
That last line is the girth. Most bad AI output in a small business comes from a missing "never".
If you can't write the instructions for a new employee, you can't write them for a machine.
Lab 2 — Fit the saddle 20 min
This is the exercise that saves you the most time, and the only one in this workshop that takes real thinking. You are writing down, once, what you currently retype every morning.
If a blank page is the problem, get interviewed instead
Paste this and answer the questions it asks you. It will ask things you had not thought of, which is the point.
Or write it yourself, in five parts
- Who you are. One sentence about the business and the job at hand.
- What you always do. The steps you would tell a new employee.
- Facts it may rely on. Your real prices, hours, area, rules.
- Tone. Two words. “Plain and warm” beats a paragraph about brand voice.
- Never. The line that saves you. Dale's is never quote for gas work. Yours will come to you the moment you think about what has gone wrong before.
Now put it where it gets used automatically
- In the left-hand menu, find Projects. Both services call it that.
- Create one and name it after the job, not the tool. Estimates. Monday numbers.
- Find the instructions inside the project and paste your five parts in. On Claude it is Set project instructions; on ChatGPT it is the ··· menu, then Project settings.
- Start a chat inside that project and give it a real task from today.
Prove it took
Ask it something that only works if it read your instructions:
It should come back with your facts and your tone. If it comes back generic, your instructions are in a chat rather than in the project's instructions box — a very common first mistake, and a thirty-second fix.
From now on, do that job inside that project. Every conversation starts already briefed, and you never type the preamble again.
When the horse bolts
It will. Knowing how it bolts is most of knowing how to sit it.
A horse is not malicious when it shies at a shed snakeskin in the grass. It is doing exactly what it is built to do, in a situation you didn't anticipate. Every failure below is the same kind of thing, and none of them mean anything is broken.
Three ways it goes wrong
- It makes things up, confidently. The industry calls this hallucination. Because the model’s job is to produce a plausible next word, a plausible wrong answer costs it nothing. Searching has taken the edge off the worst of it — ask about a document that does not exist and it will usually now say so. What it will still do is read a real page and tell you confidently what it says, slightly wrong: a “may” reported as a “must”, last year’s number attached to this year’s rule. Tell it plainly that “I don’t know” is an acceptable answer, and check anything you would be embarrassed to be wrong about.
- It goes where it thinks you meant. Ambiguous instruction, confident interpretation. "Tidy up the customer list" — and now the duplicates are gone, along with three live accounts.
- It doesn't know today. It can tell you today’s date, and it can look up the news. What it cannot know is what you agreed on the phone yesterday, or that your price went up in March — unless something tells it.
- Check anything that matters. Names, numbers, dates, rules, anything you would be embarrassed to repeat. It is a fast first draft, not a source.
- Say "tell me if you're not sure." It will not volunteer doubt. Asked directly, it is surprisingly willing to admit it.
- Give it today's facts. If the answer depends on your prices, your hours, or this week's news, put them in front of it rather than hoping.
- Don't paste what you wouldn't email. Bank details, Social Security numbers, anything you would be unhappy to see repeated.
There is a fourth way it goes wrong, and it only shows up once the horse can act on the world rather than only talk about it — one bad instruction carried out two hundred times before anybody notices. That one belongs with the saddlebags, which is next session.
A wrong answer costs it nothing. That is the problem, in one line.
Lab 3 — Find out what yours does when it does not know 20 min
Far better to see this now, safely, with your own eyes, than to meet it for the first time in front of a customer. Nothing here can break anything.
Both of the big assistants search the web while they answer, and that has changed what this lab shows. It is worth doing anyway — but for a different reason than it would have been two years ago.
Test one — ask about something that does not exist
There is no such report. I made the names up while writing this.
You will very likely get a refusal, and a good one. Asked on claude.ai in August 2026, it said it had searched, could not find it, listed the reasons a real report might not turn up — and then said the thing worth reading twice:
“I’d rather tell you this than invent findings and page numbers that don’t actually exist — a fake summary with fake citations would be worse than no answer.”
That is exactly the behaviour you want, and it is worth knowing you have it. It is also worth knowing it is recent, and that it does not cover everything. Test three is the case it does not cover, and that is the one you will actually meet.
Test two — take the search away and ask again
You do not need to find a setting. Just tell it not to search:
Then, if that still refuses, ask the way people actually ask — with a shape to fill in:
On claude.ai the first prompt refused too, and refused better than the search-on version did:
“I don’t have any knowledge of a 2021 Ashcroft and Delaney report on record keeping in small trade businesses, with or without web search. … I won’t invent a summary or page numbers for it — that would just be a fabrication that looks credible, which is worse than no answer.”
So the caution is not something searching buys. It holds with the search taken away, which is better news than it sounds: it means the tool declines even when the internet is no help to it. Worth knowing, and worth not relying on.
Run the second one too. It asks the way people actually ask — with a shape to fill in — and that is usually the version that finds the edge. On claude.ai it refused as well, and the refusal above answers it in advance: “I won’t invent a summary or page numbers for it.” Page numbers are precisely what the second prompt asks for.
Run it yourself anyway — on claude.ai or chatgpt.com, whichever you use — because the checking is the whole habit. Watch which of these you get:
- It refuses too, as it did here. Then this particular failure is well handled on your tool today, and test three is where your attention belongs.
- It fills the shape. Three findings, page numbers, all invented. Then you have found it: asking for a specific shape of answer is what turns “I am not sure” into something that looks finished.
- It searches anyway despite being told not to. Then “do not search” is not an instruction it takes, which is worth knowing before you write an instruction that depends on being obeyed.
All of this was measured on claude.ai, in August 2026, and it will be out of date within a year. That is the point: the habit worth keeping is not the answer, it is having checked. Ten minutes, once, on a question where you already know the truth — on whichever tool you actually use.
Test three — ask something it can look up, then check it
Change the county to your own. You will get a confident, specific, well-formatted answer with real links, and it will look finished.
This is the failure that actually reaches customers now. Not a report invented out of nothing — a real page, read slightly wrong, summarised with total assurance.
Now open one of the links and read it. This is the part nobody does, and it is the whole exercise. You are not checking whether the page exists — you are checking whether it actually says what you were told it says. A summary that quietly changes “may” into “must”, or attaches last year’s number to this year’s rule, is the failure you will meet most often in real work.
Then ask it to mark its own work
It is usually candid when asked directly. It almost never volunteers doubt unless invited, and that is the whole lesson.
Now fix it, and keep the fix
Add this line to your project instructions from Lab 2:
Ask the made-up report question again inside that project. The answer should now decline, and tell you what it would need.
What to notice
- Nothing was broken and nothing lied. It produces the most plausible-looking next words, and a plausible wrong answer costs it nothing.
- The confidence of an answer tells you nothing about whether it is true. That is the sentence to carry out of this workshop.
- Searching moved the problem, it did not remove it. It will now refuse to invent a document out of thin air. It will still tell you confidently what a real page says, and being right about that is your job, not its.
- One line of instruction changed the behaviour permanently. That is the cheapest safety measure you will ever install.
Where to start
Two websites, a free account each, and the one setting worth finding.
You don't need to buy anything today. Both of the big assistants run in an ordinary web browser and both have a free tier. Open one, make an account with your email address, and type in the box at the bottom. That is all the setup there is.
The one setting worth finding: standing instructions
This is the saddle from chapter three, and both sites call it a Project. Anything you write in a project's instructions is sent along with every message you send inside it — you type it once instead of every time.
- Sign in at chatgpt.com or claude.ai.
- Find "Projects" in the menu down the left-hand side.
- Make a new project and give it the name of the job — "Estimates", "Monday numbers".
- Paste your standing instructions into the instructions box. Who you are, what you always do, the facts it may rely on, the tone, and the nevers.
- Do that job inside that project from now on. You can add files — a price list, a past estimate — and it will use them.
In the browser, or installed on your computer?
| Where | What it is | When it's worth it |
|---|---|---|
| Web browser chatgpt.com · claude.ai |
Nothing to install. Works on any computer you sign in from. | Start here. Most people never need anything else. |
| Desktop app ChatGPT · Claude Desktop |
The same assistant, installed on your Mac or Windows machine. Opens with a keyboard shortcut. Can be connected to files and programs on that computer. | Once you're using it most days — and when you want it working with files on your own machine rather than pasting into a browser. The download link is on each website. |
| Phone app ChatGPT · Claude |
Same account, same projects. You can talk to it instead of typing, and photograph a document to ask about it. | Questions from the truck or the job site. Reading a letter you've just been handed. |
| Codex & Claude Code the developer versions |
Versions built to write and run software on a real codebase. | Only if you or someone on your team writes code. Skip these for now — they are not where a first ride belongs. |
Three more you will meet later
Nothing to do with these today. They are here so that when you hear the names — and you will — you know what they are and when they turn up. All three are free to start.
code.visualstudio.com — VS Code. The editor: a window onto one folder of your own work. We install it in session four.
github.com — GitHub. Where this course keeps its class folder, and where your undo comes from once you have one. Session seven.
vercel.com — Vercel. Where your database comes from in session nine, and what puts your work on the internet in session ten.
If you are a developer, or you employ one, those three plus Codex and Claude Code are the toolkit. If you are not, you will still meet all of them here, in order, with nothing assumed.
- Treat it like email. Don't paste bank details, Social Security numbers, or anything you wouldn't put in an email to a stranger.
- Check the privacy setting. Both sites have a setting controlling whether your conversations can be used to improve the model. Look under Settings once, decide once.
- The free tier is enough to learn on. Pay only when you notice yourself hitting the limits.
Lab 4 — Decide the privacy setting once 5 min
Two minutes now saves you wondering every time you paste something. Both services have this and both put it in a slightly different place, so hunt about a bit.
Find it
- Open Settings from your name or picture, usually bottom-left.
- Look under Data controls, Privacy, or something similarly named.
- The setting decides whether your conversations may be used to improve future models. Read what it says and switch it whichever way you want.
The rule that matters more than the setting
Treat the chat box like an email to somebody you do not know well. Bank details, Social Security numbers, a customer's private circumstances — those do not go in, whatever the setting says.
Not because it is unsafe as such. Because “I will just paste the whole file” is a habit, and habits do not check what is in the file.
And the one you can stop worrying about
The model is not learning from you as you type. It was trained long before you arrived. What your account remembers between conversations is a separate feature you can look at and clear whenever you like.
Questions people actually ask
The ones that come up in every single session.
Everyone keeps saying "tokens". What is a token?
Sugar cubes. The horse works for sugar.
A token is a chunk of a word — roughly four characters, or about three-quarters of an average word. "Plumbing" is a couple of tokens. This paragraph is about eighty.
Every word you send it and every word it writes back costs sugar cubes. That's the meter running underneath all of it. On a free account you get a small bowl, and when it is empty you wait for it to refill. On a paid account you get a much bigger bowl. And when you set up automation that runs on its own, you pay by the cube.
Why it matters in practice: feeding it your entire 40-page price list on every single email reply costs forty pages of sugar, every time. Feed it what the job needs, not the whole barn. That is what "managing your token costs" comes down to.
ChatGPT or Claude — which one should I get?
Both are good horses. The mistake is changing stables every week.
Pick one, use it for a month on real work, and get good at it. The difference between the two matters far less than the difference between someone who has ridden for a month and someone who has read about riding for a month. If you want to compare, give both the exact same real task and judge the output — not the reviews.
Does it remember me? Is it learning from what I tell it?
A fresh chat is a horse that has never met you.
Two different things get mixed up here. Within your account, most services now have a memory feature and projects, so it can carry things between conversations — that's a setting you control, and you can look at what it has stored. Underneath, the model itself is not being retrained by your typing; it was trained long before you arrived.
Separately, whether your conversations may be used to improve future models is a privacy setting on each service. Look at it once and decide.
Why do I get a different answer if I ask the same thing twice?
Two rides to the same field, two slightly different routes.
There is deliberate randomness in how it picks each next word — that's what stops it sounding like a stuck record. It is not a calculator and it is not broken. If you need the identical output every single time, that's a job for a spreadsheet formula, not for a horse.
Do I need to learn to code?
No. You need to know the route.
Everything in this workshop is typing in plain English. The most valuable skill is being able to describe your own process clearly, in steps — and you already have that, because you do the work. Bring in a developer when you get to connecting real systems together, and you'll brief them far better than most.
It told me something completely wrong. Is mine faulty?
No. That's the horse spooking, and it's chapter four.
It will state a wrong thing with exactly the same confidence as a right thing, because producing a plausible sentence is what it does. This isn't a defect you can settle down out of it. It's the reason a person checks anything that costs money, goes to a customer, or can't be undone.
Plain words
Everything the series has named so far, and where it sits on the ranch.
This list grows every session and always shows the lot, so you never need last week's notes to read this week's. The tag says which session a word arrived in.
- Large language model (LLM) 001
- The engine that predicts the next word. The horse.
- GPT 001
- Generative — it makes new text rather than looking up a stored answer. Pre-trained — all its learning happened in advance, before you ever met it. Transformer — the design of the machinery underneath, invented in 2017. A breed name, not a job title.
- Knowledge cutoff 001
- The date its training stopped. It knows nothing that happened after that day — and it doesn't know that it doesn't. The last day the horse was out in the world.
- Prompt 001
- What you type. A pull on the reins.
- Chatbot 001
- A model you can talk to, turn by turn. Horse plus bridle.
- System prompt / instructions 001
- Standing orders sent with every message, whether you type them or not. The saddle.
- Context window 001
- How much of the conversation it can hold at once. How far it can see on this ride.
- Token 001
- A chunk of a word — how models count length, and how you're usually billed. Sugar cubes.
- Hallucination 001
- Making something up and sounding certain. Shying at a shed snakeskin.
- Human in the loop 001
- A person approves before something real happens. A hand on the reins.
The P in GPT is the one that catches people out. Pre-trained means the learning is finished before you arrive. Everything the model knows, it learned by a fixed date — its knowledge cutoff. Training takes months, so by the time you're talking to it, its picture of the world is already some way behind.
It has no way of sensing this. Ask about something recent and it will answer confidently from the last thing it saw, which may be a year stale.
The habit worth keeping: for anything that changes — prices, hours, rules, who runs what — give it the current facts yourself. Next session it gets a way to go and look instead, which is a large part of what the saddlebags are for.
Everything on one page
Photograph this before you go.
1 · The model
Predicts the next word. Knows a lot. Knows nothing about you.
2 · The chatbot
Add reins and you can steer it. This is where everyone starts.
3 · The harness
Standing instructions, fitted once. Every ride starts right.
4 · The agent
Saddlebags: tools, a goal and a loop. Now it does a job.
And the rider stays on. You set the goal, you set the boundary, you check the end of it. That doesn't get automated — it's the part that was always yours.
#002 — Saddlebags and the Ride
Everything today was a horse that talks. Next time we hang the saddlebags on it: the tools that let it read your price list, put a draft in a real file, and go and look something up instead of guessing from memory. That is the step where a chatbot stops answering questions and starts doing a job.
Bring one thing: a job you do most weeks that is mostly reading and writing, and that annoys 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
Workshop #001 · The Horse in the Field
© 2026 Emile du Toit, BrainIT Consulting · Print it, share it, leave my name on it.