AI Workshops
All 12 workshops ↗
AI Workshops
All 12 workshops ↗
Twelve workshops in one session. The full version is there too — All twelve workshops, to work through on your own, at your own pace. Only have thirty minutes? Start with Tacking Up.
Presented by Emile du Toit
brainitconsulting.com
Rabun County Public Library
73 Jo Dotson Circle
Clayton, GA 30525
Ask at the front desk, or call
706.782.3731
Ages 12 and up.
Little or no experience is exactly right.
Ten, with a waiting list of up to five.
Bring a laptop if you have one — it is not required.
More dates go up here as the library confirms them.
New: a beginner’s guide to letting Claude or ChatGPT search Gmail
You are in the right room even if you have never written a line of code — and especially if you half suspect you have already missed the boat. We start at the beginning, we go at the pace of the room, and no question here is too basic to ask out loud.
Now, the horse.
I grew up on a farm in South Africa and rode to school on horseback. So when I say a large language model is like a horse, it is the thing I actually know rather than an image that sounded clever. I have tacked one up in the dark, been thrown off in front of people, and learned early that a horse does not care how clever you are, only how clearly you ask.
All of that is true here. It is strong, it is fast, and it is no use to you standing in a field. It needs its tack on before it can do a day's work. It does what you asked rather than what you meant. And when it throws you, it is a horse, not a mirror. So the analogy is doing real work, not decorating the slides — and where it stops being useful, I will drop it and tell you the thing plainly.
One more thing: this is meant to be fun, and none of it is dumbed down. Briefs and version control, deployments and rollbacks, checks that run without being asked — the same concepts a professional team uses on Monday morning, out of thirty years in enterprise IT and a computer science degree. You get the real thing, in words that make sense the first time you hear them.
So ask whatever you like, whenever you like — that is what the room is for. But we do have a fair bit of ground to cover.
Right. Enough horsing around — let's get started.
The shape of the session. Print it if you like knowing what is coming.
| Clock | Mins | Stop | What happens |
|---|---|---|---|
| 0:00 | 3 | Welcome | The promise, and the one rule: ask before you type. |
| 0:03 | 10 | 1 · The horse | What the thing actually is. Standing instructions. |
| 0:13 | 8 | 2 · Saddlebags | Tools and the loop. Why an agent is duller than it sounds. |
| 0:21 | 12 | 3 · The tack room | Lab A. VS Code, and one folder for one job. |
| 0:33 | 15 | 4 · The agent | Lab B. Put it in the folder. First ride. |
| 0:48 | 12 | 5 · The trail map | Lab C. Say what, never how. |
| 1:00 | 20 | 6 · Build it | Lab D. The page appears. |
| 1:20 | 12 | 7 · Breadcrumbs | Lab E. The undo, and GitHub. |
| 1:32 | 15 | 8 · To market | Lab F. Live on the internet. |
| 1:47 | 5 | The rest of the trail | What we skipped, and where it is written down. |
| 1:52 | 8 | Questions | And the one thing to go home and do. |
Watch Lids down, hands off. No account, no installs, nothing to sign up for. You watch it happen on the screen, and every step is written down on this page to take away. If you came mainly to find out whether this is for you at all, that is a perfectly good reason to be here.
Free The one I steer most people to. A free account at claude.ai or chatgpt.com, open in a browser tab beside VS Code. You describe, it writes, you paste. Costs nothing, needs no card, and reaches exactly the same live web address as the paid route.
Paid Only if you already want to build. Claude Code or Codex, about $20 a month, where the agent works in the folder itself instead of handing you text to paste. Worth it if you are going to do this every week. Not worth it if you are still deciding whether you like the idea.
My order is: free first, and pay only when the free way starts annoying you. You should never feel you need to enter a card to keep up in a free workshop at a public library, so I will say that at the start and mean it.
Do these if you get the chance. Most people will not, and the session is built so that is genuinely fine.
If none of this is done, you are still fine.
We install together at stops three and seven, and anything we run out of time for is written down right here.
Of everything on that list, the business facts are what decide how your session goes.
Come with your real phone number, your real hours, the real job you keep being asked for and turning down. Build the page out of those and it is worth keeping. Build it out of made-up details and it is a demo you will forget by Tuesday.
This is the twelve-session series compressed hard. I say it once, clearly, and then we get to the keyboards.
A large language model is a horse. Strong, fast, and no use to you standing in a field. It has no idea who you are, it forgets you between conversations, and it will confidently walk into a fence. The bridle is the words you steer it with. The saddle is the standing instructions you write once so every ride starts the same way.
Nobody blames the horse for not knowing where the feed shed is. It has never been here before. Neither has this thing, every single time you open it.
Ten minutes on what it actually is. It is good at drafting, summarising, rewriting, explaining, and first attempts at almost anything. It invents things when it does not know — confidently, and in exactly the same tone it uses when it does know. It agrees with you far too readily, so asking it “is that right?” is close to useless. And it cannot remember yesterday: every conversation starts from nothing. If you have questions, hold on to them — I answer questions at 1:52 and I will hold to that.
A chatbot talks. An agent does. The difference is the saddlebags — the ability to read a file, write a file, run a command, go and look something up. Give it a goal and let it check its own work, and that loop is all anybody means by “agent”, which is a much duller word than it sounds.
For the rest of the session it is not answering you. It is working for you, in a folder on your own laptop. Lids up.
Six stops, six labs. Each is on the screen while I talk, and I work at your pace — a step behind you, not ahead.
Work lives in folders. That is all there is to it. The one rule that keeps everything after it safe is one folder, one job — because an agent you point at a folder can see everything in it, and an agent pointed at your entire Documents directory is a horse loose in the house.
One thing to know for the rest of the session: everywhere these steps say Ctrl, you press Cmd instead.
daleandson-page, not Dale & Son's New Page (final). Spaces and apostrophes cause a specific kind of misery later.The folder step is three lines in a terminal. On Windows, open Command Prompt or PowerShell from the Start menu:
That last line — code then a space then a full stop — means “open VS Code here”. The full stop means “this folder”, and you will see it again all session.
On a Mac it is the same but mkdir instead of md. If code is not recognised, open VS Code, press Cmd+Shift+P, type shell command, and pick Install ‘code’ command in PATH. On Windows it works already, as long as you ticked Add to PATH when you installed.
If you opened Documents instead of the new folder, do File → Open Folder again. It is worth the sixty seconds to get right.
The title bar now shows the folder name. Everything from here happens inside those four walls. That is the whole reason we made it.
Nothing is installed yet — no VS Code, and no agent either. The agent arrives at stop 4. So this one goes in an ordinary browser tab: claude.ai or chatgpt.com, the free account from the bring-list. Windows or Mac, same two websites. Paste it and it walks you through at your own speed instead of mine.
On a Mac, change the first line to I am on a Mac before you send it. The rest of the prompt is the same, and you will get Finder and Cmd instead of File Explorer and Ctrl.
This is worth knowing on its own: when you are stuck on anything, asking to be walked through it one step at a time works far better than asking for the whole answer at once.
Now the agent goes into the folder with you. It can read what is there, write new files, and ask permission before it does anything it cannot take back. The permission prompt is the gate: it is not the software being timid, it is you keeping the reins.
Do this bit even if you paid. Two windows side by side is the working posture for the rest of the session, and it is worth arranging properly once.
Nothing in this session needs a paid plan. Skip this whole box if you are happy pasting — you will finish with exactly the same live website. It is here because sooner or later everyone asks what the paid version does, and the honest answer is: the pasting, by itself.
claude.Both of them have a VS Code extension, so the agent sits in a panel beside your files instead of in a separate window. If you are going to use one of these regularly, this is the version to have.
The extension and the desktop app are the same thing wearing different clothes, and they share the same sign-in. Install whichever you prefer; you do not need both.
There is also winget install Anthropic.ClaudeCode on Windows, or brew install --cask claude-code on a Mac, if either is more familiar to you.
Hold your pointer over the green button at the top left of the browser window and choose Left from the menu that appears. Then click VS Code to fill the other half.
Close that terminal, open a new one, and type claude. If you use Homebrew, brew install --cask claude-code does the same job.
Identical, except the shortcut for the Extensions panel is Cmd+Shift+X. Check the publisher says Anthropic, or OpenAI for Codex, before you install anything.
Only an agent in the folder can do this on your own laptop: Claude Code, which needs a paid plan, or Codex, which a free ChatGPT account can also run for quick tasks. On the browser route, read it rather than run it — ninety seconds, and it is the clearest picture of the difference between the two.
The answer is “nothing”, and that is the right answer. It went and looked. Then:
Watch the permission prompt when it appears. It is asking because it is about to change something on your disk, and it will keep asking until you tell it not to. That prompt is the most reassuring thing in the whole session.
The file appears in the left-hand panel. Most people say “oh” out loud the first time.
Pro tip — the file it reads before you say anything. You will get tired of explaining who you are at the start of every conversation. Put it in a plain text file at the top of the folder and the agent reads it whenever it starts work there. This is the saddle from stop 1, living in your folder instead of in a browser setting.
For Claude Code call it CLAUDE.md. For Codex call it AGENTS.md. If you use one tool, one file is all you need.
If you use both, do not keep two copies — you will edit one and forget the other. Put everything in AGENTS.md and make CLAUDE.md a single line:
Claude Code reads CLAUDE.md, and that line pulls the other file in. One file to edit, both tools reading it. Anthropic’s documentation recommends this; it also offers a symlink, but that needs administrator rights on Windows, so use the import.
Do not write it from a blank page. Get interviewed instead:
From then on, every conversation in that folder starts with your business already explained. Workshop #006 is entirely about this file, and #007 adds your own vocabulary to it.
Write it once. Every ride after that starts properly briefed.
The map is worth more than the horse. The rider’s goal is a nice campsite with a fire. Before anything gets built, you write down what you want in plain English — what it is, who it is for, and the one thing a visitor should do. Not how to build it. What, never how.
Everything from here builds your page, with your facts, your phone number, your “we don’t do gas”. Practising on something invented feels safer and teaches you less — an invented bakery never argues back, and you end up with a page you would never send to anybody.
The more ordinary the trade, the better this goes. A septic service or a fence contractor works better here than a photographer, because a page that does the job for those does the job for anything.
No business of your own? Borrow one you know well — a friend’s, or the last trade you paid for. Real facts are the point, not whose they are.
Worth thirty seconds of honesty here, because it is funny and it is true: the worst clients are programmers. They send a database schema and a list of tables when what was wanted was a paragraph about the business. They have described a solution and skipped the problem, and now everything downstream is guessing at what they actually needed. Non-technical people are better at this because they have no vocabulary to hide behind — they just say what they want.
A specification is not a technical specification. Nobody tells the builder which brand of nail to buy.
brief.md — into the folder you made in Lab A, at the top of it, not inside a sub-folder. Check the file list on the left: if brief.md is not there, it went somewhere else. Save it again and watch that panel.What the .md is, and why it does not matter. It is an ordinary text file. The # at the start of a line just marks a heading, so the assistant can tell your five sections apart. Nothing breaks if you get it wrong, and you can call the file brief.txt if you prefer — just say that name when you ask it to read the file. Save it from VS Code rather than Notepad: Notepad on Windows quietly adds .txt on the end, so you get brief.md.txt and then nothing can find it.
If a blank page is the problem, do not write the brief — get interviewed instead. Paste this and answer the questions it asks you.
It will ask you things you had not thought about. That is the point, and it is the same thing a good builder does before quoting.
On the free route nothing gets written. A browser chat cannot reach your folder, so the last line of that prompt quietly does nothing — the brief comes back in the chat window instead. Select it, copy it, then in VS Code press Ctrl+N (Cmd+N on a Mac) for a blank file, paste, then Ctrl+S and save it as brief.md in your folder. Thirty seconds, and Lab D then works the same for you as for anyone who paid.
Plenty of people do, and “build a landing page” is not much use to you if you already have one. Use this brief instead. It is the same exercise, and it produces something your existing website almost certainly does not have.
Then add the line that makes it work with no website behind it:
If you would rather not write the brief out by hand, paste this instead and answer the questions it asks you.
It will come back and ask you about your event. Answer in plain sentences. When it is done, click the button yourself before you show anybody — your mail app should open with the subject already written.
Use your real email address. A mailto link needs no database, no hosting account and nothing to maintain — it just opens the visitor’s own mail app with the subject already written. Most people assume a signup form needs something they cannot afford. It does not.
Nowhere in that file does anybody name a colour, a font, a framework or a layout. If you have written “use React”, or anything like it, take it back out and ask yourself why it was in there. The answer is almost always that you heard the word somewhere.
Give this twelve minutes and then move on, even unfinished. It is the stop people lose an hour to, because writing about your own business turns out to be enjoyable. The brief can keep growing all week; the rest of the trail cannot wait for it.
Twenty minutes and a page exists. It will not be beautiful. It will have your phone number on it, your words, your opening hours — or the date of your party and a button that actually works — and it will be yours in a way that a template from a website builder never is.
One message:
That last sentence is the one worth stealing. It hands back a list of every gap in your brief, which teaches stop 5 better than stop 5 managed to.
brief.md into the chat tab.index.html in your folder.Identical instruction, one word changed: build a single-page signup as index.html. Everything else — one file, no frameworks, works on a phone — stays the same. When it comes back, click the button: your mail app opens with the subject already filled in.
Read the instruction before you send it and guess what will come back. Then send it. If it comes back slightly wrong, so much the better — watching it get fixed in one plain sentence teaches more than watching it work first time.
index.html → Reveal in File Explorer, or Reveal in Finder on a Mac.This is the loudest moment of the session. Take ninety seconds over it — you asked for a thing in plain English and the thing exists.
Look at the address bar. It says something like file:///C:/Users/you/Documents/…. That is not a web address — it is your browser opening a file off your disk, the same way it would open a photo.
A local preview server is a small program that runs on your own machine and hands the page over properly, at an address like http://localhost:3000. Nobody else can reach it. It is your page, on your computer, pretending to be on the internet.
For one plain HTML file, honestly, it does not. Double-clicking works and you have just seen it work. It starts to matter the moment the page is more than one file:
file:// and be fine on localhost.file://. This is the one that catches people out.So: it is not a thing you need today. It is the thing to know exists, for the first time something works when you double-click it and breaks when you publish it — or the other way round.
It will pick something sensible — most machines already have one way or another of doing this. The last line is worth asking for every time: a server you cannot stop is a small annoyance you will meet at bedtime.
index.html in the file list and choose Show Preview.The page opens inside the editor at a localhost address, and it reloads itself every time you save. Once you have that, you stop double-clicking anything.
One thing that catches everyone out the first time: localhost only exists on your machine. You cannot text that address to anyone — it will not work for them, and it stops working for you the moment you stop the server. Making it real for other people is Lab F.
Whatever bothers you first, fix it in a sentence:
Reload the browser tab. And there is the thing that matters: you just described a change in the same words you would have used to a person, and the change happened. That is the skill. Everything else today is plumbing.
Now that something exists, it can be ruined. A commit is a marked point on the trail that you can always walk back to. It is the undo — and unlike Ctrl+Z it survives closing the laptop, and it goes back weeks rather than seconds.
git --version.The Source Control panel, Initialize Repository, the commit box, Discard Changes and Publish to GitHub all look and work exactly the same. The only difference is Cmd+S to save instead of Ctrl+S.
first version. Click Commit.That is a marker. Whatever happens from here on, this exact point can be got back.
index.html. Select a big chunk in the middle and delete it. Save.index.html under Changes → click the curved arrow (Discard Changes) → confirm.Forty seconds, start to finish. This is the one that stops the fear: you broke it on purpose and got it back, so the next time you break it by accident you already know what to do.
Your folder now exists in two places. If the laptop is stolen tonight, the work is not gone — which is a better reason to care about version control than anything anyone will tell you about version control.
The clicking above is worth doing once so you know what the buttons are. After that, the agent can do it for you:
The tool is called gh, and it is optional — every step above works without it. If it cannot reach GitHub, that is normal and not a failure — signing in needs your password and your own browser. Do that part with the Publish to GitHub button above.
Nobody buys from your barn. The file on the laptop is a rehearsal; the address you can text to somebody is the thing. And because GitHub already has the folder, publishing is a matter of pointing something at it.
Text the address to somebody else, right now. Nothing makes it real like a person outside opening it.
Pro tip — Hobby or Pro? Hobby is free, takes no card, and is the right choice for today. Two things to know before you rely on it.
A hundred deployments a day. Every push to GitHub is one deployment. A hundred is a great many for one page — but if you sit down for an afternoon of small fixes, pushing after each one, you can reach it. Push once when a batch of changes is finished rather than once per word. Pro raises it to 6,000.
Hobby is for personal use. Vercel's own wording: the Hobby plan “restricts users to non-commercial, personal use only.” A page for your business is commercial. Nobody is going to switch your site off for a plumber's opening hours, and Hobby is the right way to learn this today — but once it is the address on your van, Pro at $20 a month is the honest answer, and it is the plan you are meant to be on.
Checked on vercel.com/docs, 6 October 2026. Plans change — look before you trust me.
index.html. Save.fixed the hours → Commit → then Sync Changes.You do not have to open the file at all. Say what you want in plain English and let it do the editing, the marking and the sending:
That middle line is the one worth keeping. It costs you five seconds and it means nothing goes to a live page without you having seen it first.
This works whether or not you did the command-line setup below. The agent is only doing the editing and the git part; Vercel is watching GitHub and starts the deployment on its own.
And that is the shape of everything professional: from now on, saving your work is publishing your work. That loop — change, commit, live — is what a software team does all day, and you have just done it.
The clicking above is worth doing once, so that you know what Vercel actually is. After that you can hand the whole arrangement to the agent, and the point of doing so is what comes after: publishing stops being a task and becomes a sentence.
This needs the Vercel command-line tool, which needs Node.js. Ask for both at once and let it check rather than guess:
That middle sentence matters. vercel login prints a short code and then waits — you open the link, sign in, and approve the code. The agent cannot sign in as you and should not be asked to. It hands you the link; you do the approving; it carries on.
Then the arrangement itself. This is the part worth having, and it takes one message:
From that moment your entire publishing process is this, whenever you want it:
That is the whole loop, and it is genuinely what a software team does all day. The change goes out in about ten seconds, and you never open the Vercel website again unless something is wrong.
Pro tip — write it down once, in the agent’s own filing. If you are going to publish this way every week, the whole routine can become a skill: a short written instruction the agent follows whenever the job comes up, so you stop re-explaining it. There is a whole stop on this at 9, and it is optional.
There is a skill whose job is writing skills. Install it first — one line, and it comes from skills.sh:
Then describe the routine in your own words and let it do the writing:
After that, “publish my page” is the whole instruction, and the four steps and the three rules come with it every time — including on the day you have forgotten them.
This one is for later, not for today. Nothing above it needs it.
If Node turns out to be missing and you would rather not install it today, stop here without regret. The website route above produces exactly the same live site, and the automatic redeploy on every push works the same way — you set it up by clicking Import once instead of by asking for it.
Everything before this stop stands on its own. You have a live address, an undo, and a way of asking for changes in plain English. Nothing here is needed to keep any of that working.
It is for one particular rider: somebody who got to the end, enjoyed it, and now wants to build something bigger than a page — a real application other people pay to use. If that is not you, or not you yet, skip it with a clear conscience. It will still be here when it is.
Skill level — advanced
Every other stop on this page was written for somebody who had never opened a code editor. This one is not. It assumes you finished the trail, kept going afterwards, and are now comfortable with a folder, a terminal, and an agent writing files into a real project.
If that is not where you are, nothing is lost by leaving it. Come back when a one-page site stops being enough.
A skill is a set of written instructions you install once, and your agent then follows whenever the job calls for it. Not a program — a folder of markdown that teaches the agent how you want a particular kind of work done. Two of ours do most of the heavy lifting for anyone building a real application.
Both live at github.com/brainit-consulting/skills, with the full documentation. They work in Claude Code, Codex, Cursor, Gemini CLI and Copilot.
start-an-appInterviews you about what you actually want to build, then scaffolds a working full-stack Next.js app around it — database, sign-in, uploads, payments, AI, landing page and dashboard. The interview is the valuable part; the scaffold is meant to look like your app from the first commit, rather than a template you then have to fight.
If stop 5 landed — that describing the job well is most of getting what you wanted — this is the same idea with a much bigger engine behind it. It asks seven technical questions and every one has a recommended default, so “whatever you recommend” is a complete answer.
Install it:
Then open an empty folder — the app is built where you start it, and it will stop rather than scaffold over an existing project. Say what you want, or run:
security-scannerAn OWASP Top 10:2025 audit of any codebase, in any language — CWEs and detection patterns, severity scoring, and a dated report written to audit/ with the file, the line, the evidence and a fix for every finding.
It is here for one blunt reason: start-an-app checks that what it built works, never that it is safe — and it is routinely used to build apps holding a small business’s customer records. Run it once the app is real, not on the fresh scaffold: most of what the sign-in and access-control checks look for does not exist until sign-in works and there is data in the database.
Install it:
Then, in the repo you want audited:
Read the report knowing that a finding is not a breach. A fresh scaffold produces mostly Low and Info — missing headers, logging that only goes to the console. Look at the Critical and High counts first, and ignore the total.
The directory those commands come from
npx skills add is not ours. It belongs to skills.sh, an open directory of agent skills: one line installs any of them, and the same skill works in Claude Code, Codex, Cursor, Copilot, Gemini and the rest — you are not choosing a tool when you install one.
Ours are listed there alongside Anthropic’s and Vercel’s. Worth a look when you want something we have not written — there are a few thousand, and the front page ranks them by how much they are actually used.
The thing worth knowing before you install anything: a skill is written instructions, not a program. It is a folder of plain markdown. You can open it and read exactly what it will tell your agent to do, which is more than most software lets you do.
Almost every horse on this page was written by hand, in code. Shapes, coordinates, curves — a few hundred lines of it. That is why they look the way they look: flat, plain, and drawn by somebody who cannot draw.
Two are not. The racehorse at the top of this stop is one of them. Here is the other, beside the hand-written version of the same animal:
Same horse, same colours, same red tack. The one on the right took a sentence.
That is the whole argument for this stop. A skill or a tool does not make the agent smarter — it gives it a longer reach. The one on the left is the limit of what an agent can do with nothing but a text editor. The one on the right is what happens when it can also draw.
You do not need to be able to draw, and you do not need to know what a bézier curve is. You need to know that the tool exists, and to be able to say what you want. That has been the lesson since stop one; this is just the loudest example of it.
I would rather tell you what we skipped than let you leave thinking you have seen the whole thing.
There are twelve written sessions. This was six of them, at a canter. Here is what did not fit.
| What we skipped | The one-line version | Where it is written |
|---|---|---|
| The well | Where information lives when a page is not enough — a real database, and asking it questions in English. | Workshop #009 |
| Riding the fences | Automatic checks that catch the day you break something without noticing. | Workshop #011 |
| Who's at the gate | Logins, who is allowed to see what, and sending email from your own thing. | Workshop #012 |
| Rules of the barn | Standing instructions written into the folder, so the agent starts every session already briefed. | Workshops #001 and #006 |
| What it costs | Tokens, plans, and what a hobby project actually runs to. Which is usually nothing. | Workshops #001 and #010 |
About the address
The free address Vercel gives you — something like dales-plumbing.vercel.app — is a real
web address. You can put it on a business card, text it to a customer, and it will still be working in a year.
You do not need to buy a domain, and nothing in this session assumes you have one.
If you later decide you want dalesplumbing.com instead, that is a small job and costs
$11.25 a year through Vercel. Ask me and I will walk you through it.
All twelve sessions are written, free, and readable tonight.
Open the full series →
Nothing behind an email address.
One instruction, small enough that you will actually do it this week.
Before the week is out, change one true thing on the page you published. Your hours are wrong. The phone number should be bigger. There is a job you keep being asked for that is not on the list.
Ask for the change in plain English, the same way you did at stop six. Commit it. Watch it go live.
That is the loop, start to finish — once with me standing next to you, and once on your own. Everything else, the databases and the logins and the checks, is that same loop with more furniture round it.
And if the horse throws you: it is a horse, not a mirror. It did not decide you are stupid. Get back on and say the thing differently.
Everything today ran on something, and nobody has told you what. It is the least mysterious part of the whole business.
Every question you ask and every answer you get back is chopped into small pieces called tokens — roughly, a token is a short word, or a chunk of a long one. Tokens are the sugar. They are what you spend, whether you are spending money or spending a free plan’s allowance.
Which means a long rambling conversation costs more than a short clear one, and pasting your entire price list into every single message costs more than putting it in a file once and pointing at it. That is the only efficiency advice you need, and you will work it out on your own within a week anyway.
Three things and then we are genuinely finished:
It is strong, it is fast, it will work for you all day and it does not get bored. It just needs sugar.
These ten come up in every session, in roughly this order.
No. You have just built and published a website without writing a line of it, and what made that work was describing clearly what you wanted. That is a writing skill, not a coding one.
A short word, or a piece of a long one. Your question and its answer both get chopped into them, and that is the unit you are spending. A long rambling conversation costs more than a short clear one.
Either. They are close enough that the difference will not be the thing holding you back. Pick one, use it for a month, get good at asking. Switching every fortnight is what actually costs you.
Today, nothing. The free plans do everything we did. If you end up using it most days you will start hitting the free limit, and then it is about $20 a month, flat. Hosting the site you published is free on Vercel’s Hobby plan, which is meant for personal projects — once the address is your business’s, their paid plan at $20 a month is the one you are meant to be on.
What you type, and any file you point it at, goes to the company running the model — that is where the thinking happens. Paid business plans generally do not train on your material; free consumer plans sometimes do. If something is genuinely sensitive, check the setting on the plan you are using before you paste it, not after.
It asks first, every time — that is the prompt you saw at stop four. And after stop seven you have a commit, so anything it does can be walked back. Those two stops are in this session for exactly this reason.
No, and it will do it again. It predicts what a good answer looks like, so when it does not know something it produces a confident wrong one in exactly the tone it uses when it is right. Never take a number, a date or a legal fact on trust.
You could, and for one simple page it is a fair choice. The difference is that what you built today is a file you own, that you can move anywhere, that costs nothing to keep, and that nobody can put a monthly price on later.
It is very good at drafting, looking things up and tidying, and no good at deciding, judging, or being accountable to a customer. The businesses that come unstuck are the ones that hand over the second kind by mistake.
That is the most common thing anybody says to me. Almost all of it was the screen rather than the idea, and you have now been through the screen once with somebody talking you through it. Do the one small thing this week and it stops being frightening.
Everything else can wait. Each one closes with what it is on the ranch.
CLAUDE.md / AGENTS.md file in your folder. Yours to read and change. The saddle.All of this has happened to somebody. None of it needs anything clever to recover from.
| What you see | What to do |
|---|---|
| VS Code says Restricted Mode across the top | Click Manage, then Trust. It is your own empty folder, and you are the author. |
claude is not recognised | The terminal was open before the install finished. Close it, open a new one. Works every time. |
| “Claude Code requires a paid plan” | Expected — the free account does not open it. Switch to the free browser route and carry on. It reaches the same place. |
| No Source Control options in VS Code | Git was installed while VS Code was running. Fully close VS Code and reopen it. |
| Publish to GitHub does nothing | Nothing has been committed yet. There has to be one commit before there is anything to publish. |
| Your repository is not on the Vercel import list | Click Configure GitHub App on that screen and grant access to that one repository. |
| Vercel deploys, but the page is a 404 | The file must be named exactly index.html and sit at the top of the repository, not in a sub-folder. |
| The Vercel tool will not install | It needs Node.js 20 or newer. If Node is missing, install it from nodejs.org — or skip it entirely and use the website route, which gives the same result and the same automatic redeploys. |
missing_scope from the Vercel tool | Your account belongs to more than one team, so it refuses to guess which. Add --scope <your team> to the command, or ask the agent to list your teams and pick. |
| The site loads but looks old after a change | Hard refresh: Ctrl+Shift+R. Then check the deployment actually finished. |