"I'm not technical"
You have said it out loud. To a bookkeeper, to whoever set up your booking system, or to the kid at the phone shop who asked whether you wanted the backup turned on.
It is not false modesty. It is an accurate description of a skill you were never taught and have never needed, so this post isn't going to tell you that you are wrong about yourself. You are not wrong. Claude Code for non developers sounds like a contradiction, and that is worth taking apart rather than talking you out of.
A decade on the tools is a real competence: reading a coat, handling a dog nobody else will touch, and pricing a job in your head while the phone rings. It is simply not this one.
What is worth a second look is not the sentence. It is the assumption sitting underneath it, that being technical is the thing being asked of you.
So, plainly. Claude Code is Claude on your own computer, able to work with your actual files instead of only talking to you about them. It reads them, changes them, creates them, and runs commands you approve. It needs a paid Claude plan, and it is not on the free one, so that is the first thing to check rather than the last.
It was built for programmers, and what it asks you to be good at is describing a job clearly, which is a different thing from writing code. That is Claude Code for non developers in a paragraph, and the rest of this is the honest detail.
The thing it actually asks of you, and you already do it
Take Steph, a two-groomer salon in Sydney's Inner West. She is a worked example we follow through this series, and her numbers are the sort you would recognise rather than a real client's books.
On a Tuesday, Priya has to take a dog she has never met, one Steph has groomed for four years. Steph briefs her at the bath: he hates the dryer on his back end, so leave that until last and keep it low; his owner asks for a trim and means a tidy, so look at the photo from March before you start; he will lift a lip at the clippers and that is a warning, not a threat.
That took 90 seconds. And a job that has never been written down anywhere comes out of somebody else's hands looking like Steph did it.
That is the skill. Say what the job is, say what a good result looks like, say what to avoid, and say what to do when it goes sideways. It does not change shape because the thing on the other side of it is a program rather than a person. You are likely to be better at it than most programmers are, because you have spent years briefing people over dryer noise.
What is genuinely new here, and it is not the language
You have already used the chat-window version of this. You type, it answers, and then you do the rest: copy the caption into Instagram, paste the policy into a document, retype the good half of a reply into your messages.
That last part has always been yours. Claude Code is the version where it is not.
Anthropic describes it as a loop of three moves. It gathers context by reading what you have pointed it at, it takes action by editing files and running commands, and then it checks its own work by reading things back and running tests. The middle move is the entire difference. It can act, rather than only answer.
That is worth sitting with rather than skating past, because it is also why this feels like a bigger step than it gets billed as. Nothing about the language changed. What changed is that the thing now has hands.
First, the wall: Claude Code is not on the free plan
Deal with this before you spend an evening on it.
Every paid Claude plan includes Claude Code. Pro, Max, Teams, and Enterprise all have it. The free plan does not, and there is no free tier of it, no trial of it tucked inside a settings menu, and no clever way around it. If you're on Free, the step before all of this is an upgrade.
Anthropic publishes those prices in US dollars and has no Australian price list at all, which is its own small headache every time the exchange rate moves. The current figures, and the two vendors who will not quote you in the currency you actually spend, are in what AI actually costs.
Claude Code for non developers: what it actually does
It is a coding tool. Say that out loud once, so nobody feels misled later. It was built for software work, it is very good at software work, and most of what is written about it is about software work.
What it does mechanically, though, is narrower and more useful than that. It reads files, changes files, creates files, and runs commands on the computer it is installed on. Those are the same four moves whether the files are a codebase or 4,000 photos with names like IMG_4471.
Which is why the jobs it turns out to be good at in a salon are not coding jobs at all. They are the jobs made of files and repetition:
- A folder of photos that needs sorting, renaming, and dating. That is the shoebox problem.
- A year of exported appointments you have never been able to ask a question of. That is asking your business a question.
- A job you have now explained three times. Written down once, it comes back the same shape every time. That is a skill, and how to build one is the payoff post of this whole series.
- The things you would otherwise re-explain at the start of every session: your prices, your policies, the way you talk. Those live in one file it reads every time, which is what a CLAUDE.md is.
None of that asks you to write code. All of it asks you to say clearly what you want, which is the briefing Steph gave Priya at the bath.
Is Claude Code for non developers safe on your own files?
Three things about it are genuinely different from a chat window, and all three are mildly uncomfortable. Better you hear them here than find them out on a Tuesday night.
One, it lives in a text window. It runs in the terminal, and what a terminal actually is takes one line: the plain window on your computer where you type instead of click. It looks austere and unwelcoming, and that is a look rather than a warning. Anthropic also publishes a desktop app for macOS, Windows, and Linux, and a browser version at claude.ai/code for anyone on a paid plan, so a terminal is not the only place it runs.
Two, it works on real files. Not a draft in a browser tab you can close and forget. Actual folders on the actual machine, the one with your client list on it. Anthropic's own documentation is blunt that it can edit or delete a file you did not intend it to, and that getting it back means having something to get it back from.
Three, it asks first. Out of the box it asks you before it changes a file or runs a command, and nothing happens to a file until you approve it. There are settings that change how much it asks, and they matter enough that one whole post in this series does nothing but work through them: what to allow, and what never to. Read that before you touch a single one of them.
Those three sit together honestly. It is more capable than the chat window because it is closer to your things, and you don't get the first without the second.
Two lines from the never list stop being advice and start being load-bearing the moment a tool can act rather than answer:
And never let it:
The second line is most of the safety story at this level, and it is not really about AI. Keep a copy of anything you'd grieve, somewhere the program cannot reach. Do that and the worst afternoon this can give you is an irritating one instead of a serious one. The permissions post sets that habit out in full, and the honest install guide carries the way back out if you decide this isn't for you.
Is it cheating?
Now the question people think but rarely type.
In this trade it arrives specifically. Is it cheating to have AI write my captions? My cancellation policy? My reply to the four-star review with no words in it? Something about it feels like passing off work that is not yours.
That deserves a straight answer rather than a pep talk, because it is not really a question about software. It is a question about whether the work is still yours.
The test that holds up is not whose fingers were on the keys. It is whose judgement is in the finished thing.
A caption is yours because you chose the photo, you know the dog, you set the tone, and you decided the claim was true. A policy is yours because you decided what happens when somebody cancels at 8am on a Saturday, and because you are the one who holds that line at the counter. Something that drafts sentences does not take a single one of those decisions off you, and it is not able to.
Where it does become cheating is specific, and worth naming out loud. It is cheating when you send something you have not read. It is cheating when you let it settle a thing that is yours to settle: a price, a coat that should have gone to a vet instead of onto your table, an incident that needs writing up truthfully. Where AI must never go in pet care is the post on that one, and it does not soften it.
Nobody has ever asked whether an invoice was cheating because it came out of a template. The only question was ever whether you stood behind what it said. That has not moved. You still decide what is true, what to charge, and what to say to the owner whose dog you had to clip right down.
What it gets wrong, said plainly
It cannot see anything you have not shown it. Not your booking system, not your inbox, not your diary, and not what happened in the salon yesterday. It also has a cut-off date on what it learned, and it does not pick your business up by being near it.
It can be confidently wrong. Anthropic's documentation says so in as many words: confident-sounding wrong answers remain possible. Nothing in the tone shifts when it invents something, which is exactly why the 10-second check in why it gets things wrong earns its keep more at this level than it did at the last one.
And mistakes on files are recoverable, mostly. It keeps checkpoints you can rewind to, and a change is reversible if there is a backup or a version history behind it. Anything that has already left your computer is a different category: a message that sent, a card that charged, a page that published. None of that rewinds, for anybody, which is why the permission question is not paperwork.
Where the series goes from here
This is the point where the ladder changes character. Below it, the tool is a window you talk into. From here, it does jobs on your things. The complete guide lays out all four levels if you want the shape of it before you commit an evening.
You are also allowed to stop. A reader who stays in the chat window and gets genuinely good at prompts has taken most of the ordinary wins already, and there's no shame in parking here. The reason to keep climbing is repetition. If you're doing the same job by hand more than once a fortnight, this is the level where it stops being your job.
The next post is the only one in this series with install steps in it, and it asks for 20 minutes.
What this actually reclaims for you
Honestly, this post reclaims nothing. It is a decision, not a job.
So here is the arithmetic on the decision instead. Reading this took about nine minutes. The install guide asks for 20. Call the whole thing half an hour, spent once.
Steph's admin evening is two hours, twice a week. Say the first month of this gives back one of those four hours and never improves after that. Half an hour, once, against roughly 50 hours across a working year. That is a worked example rather than a promise, and it is deliberately the pessimistic version of it.
The optimistic version is the rest of this level. Next: the honest install guide, 20 minutes, Mac and Windows.