What an AI coding agent actually is
It is not a chatbot that writes code for you to paste somewhere. It is a program that sits on your computer, with permission to open your files, change them, and run things — while you watch.
The difference matters. A chatbot gives you a wall of code and leaves you to work out where it goes. An agent opens the folder, creates the files, installs what's needed, starts the app, looks at the result in a browser, notices the button is broken, and fixes it. You supervise; it does the typing.
That's why you don't need to code. You need to be able to say what's wrong — "the times are showing four hours early" — which is a thing any owner can see.
Think of it as a very fast junior who has read every manual, never gets bored, and will happily redo something eleven times — but who has never met your customers and needs you to explain the business.
The agentic loop
Everything that follows makes more sense once you've seen the loop the agent runs in.
Two things follow from this, and they're the whole skill of working with an agent:
- Say what's wrong, not how to fix it. "The owner's phone number should be on the card" beats trying to describe code you can't see.
- Let it finish the loop. It will often write something, run it, find its own mistake and correct it. Interrupting at the first sign of activity is like grabbing the wheel.
The agent asks before anything that costs money, publishes something, or creates an account. If you ever see it about to do one of those unasked, stop it — that's not how this skill is meant to behave.
Install your agent
Two good options. Either works with this skill. If you've no strong feeling, take Claude Code.
First: Node.js (both need it)
Node.js is the engine these tools run on. Go to nodejs.org, download the version marked LTS, and click through the installer. Nothing to configure.
To check it worked, open a terminal — Terminal on a Mac, PowerShell on Windows — and type:
A number like v22.11.0 means you're set. An error means the install didn't take; run it again.
Claude Code
There's a desktop app for Mac and Windows at claude.com/claude-code — download, sign in, done. If you'd rather work in a terminal:
Then start it inside whatever folder you want to work in:
Codex
OpenAI's equivalent. Same idea, same skill works:
These tools move quickly and install commands do occasionally change. The official pages are always right; this page might be a month behind. If a command errors, check the vendor's install page before assuming you've done something wrong.
Two free accounts worth having
Neither is required to start. Both make what you build considerably more useful, and both are free for what you'll be doing.
GitHub
A safe copy of your work, with a full history of every change. Like version history in Google Docs, but for your whole project — including the ability to go back to how it was on Tuesday.
"Put this on GitHub for me — private."Vercel
Where your app lives when you want it online, and where its database comes from. The free "Hobby" plan covers a small business app comfortably.
"Put it online so I can see it on my phone."Sign up at github.com and vercel.com. You do this part yourself — an agent should never create an account or enter a password on your behalf, and this one won't.
Why the database matters more than it sounds
Your app needs somewhere to keep its information — the bookings, the customers, the notes. With a Vercel account, the skill sets up a free one and does something worth understanding:
Without a Vercel account it still works — the database just runs on your own machine instead. You're never blocked, you simply get less.
Install the BrainIT skills
A "skill" is a set of instructions your agent reads before it starts, so it knows how to do a particular job properly. BrainIT's start-an-app skill knows how to guide you from an idea to a working app.
Once the app is real, add BrainIT's security check:
Both skills are free and open source. They work in Claude Code, Codex, Cursor, Gemini CLI, Copilot and a dozen others.
Building your first app
The one rule that matters
Open an empty folder. The app is created wherever you start the agent — the folder you open is the folder it lands in. There's no "where shall I put this?" question, because an agent can't reliably write outside where it started.
Then just talk
- Start the skill Type
/start-an-app, or simply say what you want — "I want a booking system for my salon" works just as well. - Describe the business Not the software. What happens in your day, who rings you, what you write down.
- Answer the questions Seven of them, each with a recommendation. "Whatever you recommend" is a complete answer to any of them.
- Check the plan It reads the whole thing back — what it will remember, what you'll be able to do, and what's deliberately not in version one. Correct anything that's wrong. This is the cheapest moment to change your mind.
- Let it build Ten to twenty minutes. It'll narrate as it goes.
- Use it Then say what's wrong in your own words, and it fixes it.
When it asks what you do today — a spreadsheet, a paper diary, a WhatsApp group — answer properly, and offer to show it. The columns in your spreadsheet are, almost exactly, what the app needs to remember. That one answer saves more back-and-forth than anything else in the conversation.
How to answer well
The questions are simple, but vague answers produce a vague app. Here's the difference, using a dog-grooming salon as the example.
Opening questionWhat are you building?
Why: the second one contains the entire data model. Every noun you say — owner, dog, groom, time — becomes something the app remembers.
Question 1Who's it for — your customers, your staff, or just you?
Why: this single answer decides the kind of database, whether there are logins, and who can see what. "Everyone" leaves all three unanswered.
Question 2Do people need their own login?
Why: the reason shapes the result. Individual logins mean you can remove one person without changing everybody's password — worth saying out loud.
Question 3Will people upload anything?
Why: "maybe later" builds nothing and clutters the conversation. Saying no is a good answer — anything can be added later, and a smaller first version is a better first version.
Question 4Will customers pay you through this?
Why: payments drag in accounts, tax handling and a provider signup. Naming it as a later thing keeps your first version small enough to actually finish.
Question 5Should it do anything with AI?
Why: AI added for its own sake is a gimmick you'll switch off. AI pointed at a job you actually repeat is the one you'll keep.
Question 6What should a first-time visitor see?
Why: this decides whether you get a public page describing your business, or a locked door. A staff tool with a marketing homepage is a very common way to make something feel fake.
Question 7Do you already have a website?
Why: with an address it reads your real colours and fonts off the page and uses them throughout. Without one, a description of a look you like is nearly as good. Either beats the default grey.
Two answers that are always fine
"Whatever you recommend." Every technical question has a recommended default. You are not expected to have an opinion about databases.
"Not in version one." The skill asks directly what to leave out, and this is the answer that gets you something finished instead of something half-built.
Things worth building
The best first app replaces a spreadsheet you already maintain, or a notebook you'd be lost without. Start there, not with the grand idea.
Salon or grooming diary
Appointments, clients, and the details you only learn the hard way — hates the dryer, always late, prefers the small room.
"A diary for the salon that remembers the dogs."Vet or dental client records
Patients, owners, visit history and what's due next. The reminder list writes itself.
"Client records and what each animal is due for."Trades job sheets
Jobs, addresses, what was quoted, what was done, photos before and after. Ends the glovebox-full-of-receipts problem.
"Job sheets I can fill in on my phone at the van."Tutor or small school roster
Students, sessions, attendance and a note per lesson. Parents' reports come out of the notes you already made.
"Who I taught, when, and what we covered."Cleaning or maintenance rounds
Properties, visits, checklists and sign-off. Staff tick it off on a phone; you see it from the office.
"A checklist per property my cleaners tick off."Committee or HOA requests
Maintenance requests, who reported it, what was decided, what it cost. An audit trail that isn't an email thread.
"Track maintenance requests and what we decided."Coach or therapist client log
Clients, sessions, plans and check-ins. Private per client, with a history you can actually search.
"A private log of each client and their programme."Stock or supplier orders
What you carry, what's low, who you order it from and what you paid last time.
"What's running low and who I buy it from."If you can describe the whole thing in two sentences without saying "and also", it's the right size for a first app. Everything else is version two — and version two is much easier once version one exists.
What happens after it's built
| What | Where it is |
|---|---|
| Your app | Running on your machine. Start it any time with pnpm dev. |
| A help guide | If the app is for staff or customers, a ? in the corner opens a guide written from your own answers. New staff read that instead of asking you. |
| DESIGN.md | A plain file listing your colours and fonts. Change a colour there and the whole app follows. |
| Your data | In a real database, with a full record of every change to its structure. |
Two offers at the end
Both are offered, never done for you:
- Put it on GitHub — a backup with history. It'll be private by default, and it will tell you so before creating it.
- Put it online — a real web address you can open on your phone. Worth knowing: anyone with that link can open it, so it's a preview, not a launch.
A proper launch — your own domain name, taking real payments, appearing in Google — is a separate conversation. Getting told that plainly is more useful than being half-launched by accident.
Then check it's safe
Once there's real data in it, run the other skill:
It writes a report of anything worth fixing. A finding is not a break-in. A brand-new app produces mostly low-priority notes — read the Critical and High counts first and ignore the noise.
When something goes wrong
Every one of these came up while building a real app with this skill. Knowing the words helps you say the right thing to your agent.
"Invalid origin" when you try to sign in
The app moved to a different door number than it expected. Say: "the app is on a different port — update the auth URL and restart it."
"name can no longer contain capital letters"
Your folder has a capital letter in it, and the naming rules don't allow that. The skill handles this itself now, but if you see it: "work around the folder name."
It asks which Vercel team to use
You belong to more than one. Pick your own personal one — never a client's or an employer's, or your practice app ends up in their account.
The times are wrong by a few hours
A timezone problem, and a common one. Just say it plainly: "the times are showing four hours early." That's enough to fix it.
It's been quiet for a while
Installing takes a few minutes and looks like nothing is happening. Let it finish. If it's genuinely stuck, say "what are you waiting on?"
Describe the symptom, not the cure. You can see the app; the agent can read the code. Between the two of you that's everything — and you're the only one who knows what it's supposed to do.
If you'd rather not do it alone
Plenty of people read a guide like this, get halfway, and decide their time is better spent running their business. That's a perfectly sensible conclusion.
I'm Emile du Toit. I build software for small businesses, and I teach owners to do it themselves when that's what they'd prefer — through BrainIT Consulting.
Have it built for you
You describe the business, I build the app — the same way this guide describes, just without you at the keyboard. You end up owning the code and the accounts, not renting a subscription.
"Just build it for me."Learn to build it yourself
Sitting alongside you while you build your own first app, so the second one doesn't need anybody. Best if you want this to become a skill rather than a purchase.
"Walk me through my first one."Get unstuck
You've built something with an agent and hit a wall — it works but you don't trust it, or it broke and you can't tell why. A second pair of eyes on what you already have.
"It works, but should I trust it?"brainitconsulting.com — have a look, or just get in touch and describe the problem.
Nothing in this guide needs me. The skills are free, open source, and yours to use — that's the whole point of writing it down. The offer above is for people who'd rather buy the time back.