Changing Your Own Website With AI: What Has to Be in Place First
By Daniel Mashkov, AI Systems ArchitectPublished 9 min read
TL;DR
Yes, a business owner can now ask an AI coding agent such as Claude Code to change her own site, in plain words, and get a working change. It is safe only when five things exist first: the code in a repository she owns, written rules the agent reads, tests that run on every change, locked areas for payments and data, and a yes from her before anything goes live. Without them, it is a faster way to break a site.
What changed: the owner can ask, and the site changes
Until recently, a business owner who wanted a new section on her home page, a holiday look for the shop or a fix to a broken filter wrote to her developer and waited. An AI coding agent changes that. Claude Code, for example, reads the site’s code, makes the change, runs the tests and shows the result, from a request written in Hebrew or English. It runs in a terminal, in the Claude app on a laptop and in a browser on a phone.
The agent is fast and usually right. Usually is the problem. A site that takes payments, keeps customer data and ranks in Google has parts where one wrong line costs money or traffic. The question is not whether the AI can write the change. It is what stands between its change and the live site.
Two kinds of change, two different tools
Most of what a business owner changes is content, and content does not need AI at all. Prices, products, photos, texts, opening hours: these belong in an admin screen, saved to a database, the way they always have. An AI agent is for changes to the site itself: a new page, a new filter, a different layout, a seasonal look, a bug. Mixing the two is the first mistake. If the owner has to ask an AI to change a price, the site is missing an admin, not an agent.
- Content, in the admin: products and prices, photos, page texts, opening hours, coupons, orders and reviews. Saved at once, no code involved.
- Changes to the site, through the agent: a new page or section, a filter, a quiz question, a layout or colour change, a seasonal theme, a bug fix.
- Neither, without a developer: payments, the database structure, the accounts (hosting, domain, payment provider) and anything that moves or deletes customer data.
The five things that have to exist first
None of these is new. Software teams have used them for years to let many people change one codebase safely. What is new is that the person on the other end of the agent is the business owner, so each one has to work without her knowing it is there.
- The code in a repository she owns, on GitHub or similar. Every change is a commit that can be read, compared and undone, and nothing is lost when a developer leaves.
- Written rules the agent reads before it starts. Claude Code reads a CLAUDE.md file in the project: how the site is built, the design rules, both languages, what never to touch. It is the handover note a new developer would get.
- Tests that run on every change and fail when something breaks: the pages load in every language, the cart and the checkout work, the layout holds on a phone. A test that cannot fail protects nothing, so each one is broken on purpose once to prove it.
- Locked areas. The payment code and the database structure sit in folders the agent may not edit. In Claude Code a hook can refuse the edit outright; the agent stops, explains why, and leaves a task for the developer.
- Her yes, then the checks, then live. The agent shows screenshots of the change and asks before publishing. After her yes, the full test suite runs again on a server, and only a change that passes reaches the live site. Any version can be rolled back.
How it works on a live store
I built this setup for Nataly Levin, a jewellery artist whose store went live on natalylevin.com in November 2026. Her pieces, prices, texts and studio hours are in her admin. For anything else she opens Claude on her phone or her laptop, connected to her own repository, and writes what she wants, in Hebrew. Her repository carries the rules for her Claude: the design, both languages, her brand’s wording, and a set of skills for the jobs she asks for most, such as dressing the site for a holiday or adding a stone guide.
Her Claude makes the change, checks it in a browser, shows her screenshots and asks whether to publish. On her yes it pushes the change with a marker in the commit message. A workflow in her repository then puts the change on top of the live code and runs every check the main pipeline runs — the type check, the unit tests, the build and the browser tests in both languages, on desktop and on an iPhone engine. Only if all of them pass does the change reach her main branch, the one her store is published from. When we rehearsed it on her real repository, the first run stopped at a failed security check and her live code stayed untouched. That is the behaviour you want.
The payment code and the database structure are locked. If she asks for something there, her Claude stops, explains in plain Hebrew why it needs me, and opens a task labelled for me. Every account is in her name, so nothing depends on my machine or my login.
Does it work on WordPress or Shopify?
Partly, and the reason is where the site lives. On a custom build, the whole site is code in a repository, so every safeguard above applies to every change. On Shopify, a theme can be connected to a GitHub repository, so theme changes can go through the same route; the store’s products, apps and checkout stay in Shopify’s admin. On WordPress, most of the page content and much of the settings live in the database, often inside a page builder, not in files an agent can change and test. A custom theme or plugin can be worked on this way; the content around it cannot. This example is a custom store, but the same rules apply wherever the code lives in a repository.
What it costs to run, and what it does not replace
The running parts are a Claude subscription for the owner, minutes on GitHub Actions for the checks, and the hosting the site already needs. The cost that matters is the setup: the tests, the rules and the locks are written once, by someone who knows the site, and kept up to date when the site changes. That is the work a developer does instead of every small change.
It does not replace judgment. The tests catch what they test and nothing else, so a change can pass and still read badly or look wrong on a screen nobody checked. The owner still looks at the screenshots before she says yes. Payments, data, accounts and anything new in those areas stay with a developer. What she gains is the large middle: the dozens of small changes that used to wait in someone else’s inbox.
FAQ
Can a non-technical business owner really change her website with AI?
Yes, if the site is set up for it. She writes the request in plain words, the agent makes the change and shows screenshots, and she approves it. The tests, the rules and the locked areas do the technical part, so she does not need to read code.
What stops the AI from breaking the checkout?
Two things. The payment code is in a folder the agent is not allowed to edit, and a test buys a product end to end on every change, with a stand-in for the payment provider. A change that breaks the purchase fails that test and never reaches the live site.
Do I still need a developer?
For the setup, for payments, data and accounts, and for anything the tests do not cover yet. Not for the everyday changes to pages, layout and seasonal looks once the site is ready for them.
Can a change made by AI be undone?
Yes. Each change is a separate version in the repository, and hosts such as Cloudflare Workers can roll the live site back to a previous version. Undoing a change is one step, not a rebuild.
Sources
Related services
Related case study
Related reading
April 2026 · 6 min
Will AI Replace Web Developers? What Owners Need to Know
September 2026 · 8 min
AI Agents for Business: What They Are, When They Pay
October 2026 · 13 min
WordPress Site Maintenance: What a Plan Must Include
October 2026 · 11 min
Shopify to WooCommerce Migration (and Back): What Transfers