Episode 3 · Resources
Claude Code and Codex, Explained From Zero
Everything from the video in one place: what these agents actually are, how they work inside, what to trust them with, the two ways a business can use one, how to set one up step by step, and the exact instructions to try both today.
Links
Use the official pages, so you're always installing the real thing.
What Claude Code and Codex are
They are close cousins. Everything here applies to both, so use whichever one your company already pays for.
Can talk
You type, it types back. Ask which customers haven't paid and the best it can do is tell you how you might find out. It can't go and look.
Can use the tools it's given
The same AI with two things bolted on: hands (open a file, search an inbox, read a spreadsheet, run a program) and a loop that keeps going until the job is done.
Can make new tools
Code is just written instructions for a computer. An agent that can write code and run it isn't stuck with the tools somebody handed it. You never read the code: you say what you want in plain English and check what comes out.
On your computer, in one folder
You point it at a folder. Inside it, the agent can open your files, create new ones, run programs, look things up and reach the systems you connect. It comes as a terminal app, a desktop app with buttons, and a cloud version.
How it works inside
Look, think, act, check
You give it a goal and it works out the steps, going around until it's sure. When it has numbers to add up, it writes a few lines of code so the computer does the arithmetic instead of the AI guessing.
The desk (the context window)
Everything it's working on has to fit on one desk: your request, the files it opened, its own notes. A crowded desk makes for a sloppy agent, so: one job, one session. When the job changes, start fresh.
The notes file
It remembers by writing things down in a plain file it reads at the start of every session: CLAUDE.md for Claude Code, AGENTS.md for Codex. Tell it "remember, we never discount more than ten percent" and it adds a line.
Skills
Save a whole job as a routine (how we close out the month, how we write a quote) and ask for it by name next time.
What it's allowed to do
You set the limits
It stays in the folder you gave it. You decide what it may do without asking and what it has to stop and ask about first: yes, no, or "yes, and don't ask me about that one again."
Treat it like a new hire
Start read-only: it can look at the inbox but not send, read the books but not change them. Only connect accounts your company has approved, and give it more as it earns it.
How it reaches your systems
Four ways, from easiest to most hands-on.
A plug made for AI
Gmail, Outlook, Google Calendar, Google Drive and Slack have one. Ask the agent to connect, and sign in.
A program you control by typing
Some software companies make one for their product, and these agents are very good at using them.
A drive-through window for software
Most business systems have one. Get a key from the company that makes yours; the agent reads their instructions and builds the connection itself.
The fallback
Just about every system will give you a spreadsheet. Save it in the folder and the agent can read it today.
Where it goes wrong, and the fixes
It guesses
Without the facts, it would rather hand you an answer than say "I don't know." So you get a price that sounds right and isn't.
It's missing something
It can't quote from a price list it can't see, and it doesn't know a rule you never told it.
It says "done" too early
The work looks finished, it sounds confident, and one case never got checked.
The fixes
Give it the real source. Make it show where every number came from. Ask it to test its own work and tell you what it didn't check. Keep a person approving anything that goes to a customer or moves money.
Two ways to use it
Ask one question: does this job happen the same way, over and over?
It is the tool
Hand it a job and it does the job now. Right when the job is a one-off, different every time, or takes reading and judgment. Nothing to set up.
It builds the tool
Describe a tool you wish you had and it builds it. From then on the tool does the job: nothing inside it is guessing, it does the same thing every time, and it costs next to nothing to run.
The ladder: do the job with the agent once and get it right; save it as a skill; then, when you run it every week and it never changes, have the agent turn it into a tool that runs by itself.
Use one: jobs to hand it
Each one has a goal, a source to work from, a rule or two, and you looking at the result.
Here are two hundred supplier invoices as PDF files. Make me one spreadsheet with the supplier, the date, the amount and the due date. Flag anything that looks like a duplicate.
Here's last quarter's sales export. Which products made money, which ones lost money, and what changed from the quarter before?
Which jobs did we finish last month and never invoice?
Look up these twenty competitors. Build me a table of what each one charges, with a link to where you found it.
Use our price list and our last ten quotes to draft a quote for this request. Leave the final number for me.
This customer list has the same people in it twice. Merge them, and show me the ones you weren't sure about.
Use two: tools to have it build
A quote calculator
A simple page your staff opens. Pick the job type, enter the size, and it shows the price, straight off your price list.
A request form
The customer fills it in, the request lands in your spreadsheet, and the customer gets a confirmation.
A morning dashboard
One page: today's jobs, this week's money in, and who's past due, pulled fresh every morning.
A night job
Every night it pulls the day's orders from your online store and updates the stock sheet. Nobody presses anything.
A bridge between two systems
A job gets marked finished in the scheduling system, and the invoice shows up in the accounting system.
Changing it later
Go back to the agent and say what you want: "Add a field for the delivery date."
When to get help
A tool for yourself is one thing: if it's wrong, you notice. Once a tool takes payments, holds customer information, has logins, or your business leans on it every day, it needs what all real software needs.
What it needs
Testing. Security. Backups. A place to run. And somebody who gets the call when it breaks.
Who is responsible
The agent can do a lot of that work. It can't be the one responsible for it. For those tools, bring in somebody who builds software for a living; they'll be using these same agents and will know what to check.
Set it up in 5 steps
Take it slow. About ten minutes the first time. No coding. These steps use Claude Code; Codex is set up in much the same way with a ChatGPT account.
Make a Claude account
Go to claude.ai. Claude Code needs a paid plan; Pro is the cheapest one that includes it.
Open your terminal
Mac: hold ⌘ Command, press Space, type
terminal, press Enter.
Windows: click Start, typePowerShell, press Enter.A terminal is just a window where you type instead of click. You can't break anything by opening it. If you'd rather skip it, both tools also have a desktop app.
Install Claude Code
On the Claude Code page, click Copy next to the install line. Click into your terminal, paste, and press Enter. Give it a minute.
You don't need to understand that line. All it does is download Claude Code and set it up.
Pick your folder
Make a folder on your desktop called
business. In the terminal, typecdand a space, drag the folder into the terminal window, and press Enter. That tells the terminal "work in this folder."Type
claudePress Enter. The first time, your browser opens so you can sign in. When it asks whether you trust this folder, say yes. From here on, it's plain English.
Try both today
First export your unpaid invoices from your accounting software and save the file in the business folder.
Read the invoice export in this folder. List every customer who is more than thirty days late, how much each one owes, and the total. Then write a polite reminder for each one, as a draft, and don't send anything.
Build me a simple page I can open on this computer. It should read that same invoice export and show who is late, sorted by the amount they owe, with a total at the top. Tell me how to open it, and how to refresh it when I save a new export.
Read what comes back, fix what's off, and tell it what you fixed. That's how it gets better. Next month, save a fresh export, refresh the page the way it showed you, and it's up to date.
Chapters
- Claude Code is a terrible name
- What is AI, really?
- From chatbot to agent
- What Claude Code and Codex are
- Why "code" is in the name
- A job, start to finish
- What it remembers
- What it's allowed to do
- How it reaches your systems
- Where it goes wrong
- Two ways to use it
- Use one: it is the tool
- Use two: it builds the tool
- Which one do you need?
- When to get help
- What is a terminal?
- Setting it up, step by step
- Your first job, and your first tool
- Recap
Want these tools built in for you?
Benmore's forward deployed engineers build inside your business: your systems connected, the tools tested, secured and looked after.
Talk to Benmore