AI is everywhere — and so is everyone telling you to use it for everything. But when it comes to your website, your plugins, and the work you ship for clients, the wrong AI shortcut can quietly become a serious liability.

Check out my Sparks Group Training for anyone and everyone involved in the publishing of websites, from copywriters to SEOs: https://websitesmadesimple.co.uk/sparks-group-mentoring/ It’s a supportive group mentoring session for website designers and developers to bring their questions, roadblocks, ideas, and “is it just me?” moments.

Welcome to a special bonus episode of Websites Made Simple with me, Holly Christie. (Quick note — my podcast mic was misbehaving for this one, so the audio is not quite up to our usual standard. My producer Neal at Podknows has worked his magic to clean it up, so please do stick with it. The conversation is genuinely worth your time.)

I am joined by senior developer Rob Fisher to talk about where AI actually helps in web design, where it absolutely does not, and the security and stability risks nobody is being loud enough about.

We get into:

Why vibe coding is not a shortcut, it is a stack of technical debt waiting to break

How LLMs really work, and why they hallucinate when you let them

Why your prompt boundaries matter as much as your prompt

The chatbot that gave a customer 85 percent off a 10,000 dollar order

What the October 2025 AWS outage tells us about AI first dev cultures

Why AI tools are junior developers, not senior ones, and what that means for your business

Why vibe coded WordPress plugins are a house of cards

How a “safe” plugin can become a Trojan horse with one auto update

What is actually changing in WordPress 7

If you are a designer or developer trying to work out where AI fits into your business without putting your clients at risk, this one is for you.

Come and find out more at https://websitesmadesimple.co.uk or email me at [email protected]

Skip to the good bits:

00:00 Bonus episode intro and audio note

01:18 Where AI fits in running a web business

02:34 Vibe coding and why it is a shortcut too far

05:12 The flip side, vibe coding as a learning tool

06:48 Tech debt: your vibe coded site is a liability waiting to break

07:45 What AI cannot do for user journeys and design psychology

09:20 How LLMs actually work, probability engines not brains

11:30 Hallucinations and the suspiciously short prompt

13:05 Prompt boundaries and staying on topic

15:10 What system prompts at OpenAI and Anthropic actually look like

17:00 The chatbot that gave away an 85 percent discount on a 10,000 dollar order

19:40 AI first dev houses and the October 2025 AWS outage

22:15 Microsoft layoffs and the AI replacement myth

24:30 Why generative AI is a junior developer, not a senior one

26:20 Vibe coded plugins, the house of cards problem

28:50 The WordPress plugin ecosystem and what it lets in

31:00 How a safe plugin becomes a Trojan horse overnight

33:15 WP Rocket, All in One Migration and the malicious update problem

34:50 WordPress 7 and the new AI APIs, and why it will not feel revolutionary

37:00 Block editor improvements and the abstraction layer problem

38:40 Final thoughts and how to get in touch

Mentioned in this episode:

Remember to sign up for our blogging masterclass!

Go to https://websitesmadesimple.co.uk/masterclass

Blogging masterclass

Hello, and welcome to this special bonus episode

of Websites Made Simple. In this episode, I am going to

talk to senior developer Rob

Fisher, and we are tackling an

area of AI. Now, I know that AI is

everywhere and no one needs a whole other podcast about it,

but this coincides with some

different ways that AI are being used within websites and website

technology. And the conversation is so

interesting, I can't wait to share it with you. There

is something to know that my podcast microphone was

not doing its thing during this recording. And

you all know I have an absolutely brilliant podcast producer, Neal Velio of

Podknows Podcasting. He has done absolutely everything he

can to clean this up and to make this a really special

episode. So it might not be the quality that you are used to, but

please do stick it out. You're going to learn some amazing things from Rob

and some great anecdotes.

Obviously, AI is playing such a huge role in

everything these days. You know, with so many AI tools

in every area of running business, whether it's setting up email

funnels, whether it's websites, whether it's putting your iPhone camera

on and the filters coming up, it is absolutely everywhere.

And in some ways that's really great because it's kind of making

our productivity really kind of efficient.

And in other places, it's not quite all

it's made up to be. What I'd really like to focus on is where the

role it plays in websites. I wanted to start off because I have an

episode about Vibe coding and how much Vibe coding absolutely

sucks. It is not my bag at all. And I

always feel like with Vibe coding, the people who are doing it are the people

who have absolutely no business in doing it as well. You know, people

who are like VAs or they're just not working in that tech space

whatsoever. Give me your thoughts on Vibe coding, Rob. Vibe

coding is an interesting phenomenon. Obviously, as a term, it's been around for a couple

of years now. To me, the concept of Vibe coding

in its most literal sense is quite terrifying because you're

essentially allowing people who don't know necessarily what

to expect or what issues to look out for

to build potentially complex software.

And as much as I've. I've always been an evangelist for

bringing people into the coding space and learning how to

code, learning how computer systems and servers work.

Vibe coding is kind of a. A shortcut. It's sort

of jumping up 10 steps and trying to operate as a senior

developer without knowing the things to look

out for, the red flags, the potential

security flaws that you could introduce simply by

not reminding your prompt to, to include them,

or perhaps the prompt just hallucinates them out.

If you don't have experience in what a software

solution looks like, you don't know what to look for at the end to make

sure it's finished and even how to test it. I think that

is the core of the problem with Vibe coding.

Yeah. And I think as software updates,

you have lots of people who say, oh, I've just Vibe Coded my website and

I've done this, that and the other. But software doesn't stay the

same. We're always kind of patching and updating and things like that.

And I think often I find people who are

saying, oh, I'll just vibe code it. And I often hear people like Vas saying

that, people who are like, oh, I work in Google Workspace and I'm just going

to get Gemini to code this or code that. And you think, God, when this

falls down or when this goes wrong or, you know, when this gets attacked or

whatever, you're just not going to know what to do. And

I, yeah, I personally feel like Vibe coding is not the

right shortcut if you want that really solid structural result.

Yeah. And I mean, on the flip side, you can use

the Vibe coding method as a way of learning.

And I know a lot of younger aspiring coders who are now doing

this. It's a new way into the industry

where they have a companion with them, the AI tool

Claude or OpenAI or whatever, and they're asking it to teach

them and tell them how things work, why they work, and

then how they can actually test that themselves. And I think that

is a, that is a valid use of

generative AI and the ecosystem around that. So I

think, I think it does have a place in the teaching environment and

obviously it has a, it has a place as a tool in a

software developer's toolkit. Generative AI can save

a lot of time for a developer who knows what they're doing.

The problem is, for a developer that doesn't know what they're doing,

that thing is flipped on its head where actually what you're doing is you're stacking

up what we call technical debt. So your Vibe

Coded software is a liability in

your platform just waiting to break. And, yeah, that can

cost a lot of time, money and stress further down the line.

Yeah, I can imagine. I mean, I obviously approach, as a website

designer, I approach things always from a design point of view and,

you know, coming across more and more Vibe Coded websites. And as

you know, I won an award for my user journeys, my work with user

journeys and stuff. And that's what

vibe coding never seems to do, is give a really good user journey.

It doesn't have that psychology behind it. And unless you're

putting the prompts in there, One of the things that we look at

building in once we've done our hero section and stuff, is looking at the kind

of Z layout, because we know how people read websites. You know, they go

from left to right, up to down, that sort of thing. And it's all those

things, it's like the. You don't know what you don't know. And often,

like you say, with putting in the prompts, I find there's even only

so far that people go with their prompts, you know, especially if they're not a

web designer who's vibe coding a web design. And I

like that. When we spoke about this before that you were saying about,

you know, ultimately AI just doesn't have a brain, so it can't kind of take

it from that psychology point of view. That's right. So

essentially your. The LLMs, the

generative AI systems that we've got at the moment,

do one job, and that's language processing

on an insane scale. And what that means is they

are essentially probability engines that guess the next word.

I mean, it's stupidly technical. There's a lot of maths behind it that I don't

understand. I read the crib notes,

I like to read them out to you. So there's a lot of, there's a

lot of stuff in there. Essentially you're dealing with a black box that is

full of words, phrases, even, to a certain

degree, meanings. But they're all connected together by chains

of probability over lots of levels. What they're

able to do, because they've been trained on the Internet,

they've been trained on human language that's been put together over

the last few thousand years. They're able to pretty

accurately produce strings of

text, strings of language that make sense in a given

scenario, which is why you're able to give it a prompt

of a certain amount of words, and it will give you an

answer that looks like, for all intents and purposes, that

it was written by a human. It's very, very clever

technology. I mean, we must remember that this is, this is

cutting edge stuff. But as you say,

it doesn't have a brain. It is literally a mathematical

equation that's being called many times over every

time you give it a prompt. And the law of

probability Means that most of the time it gives you something

that's relevant or useful, but there's always that

there's a non zero chance it's going to give you what they call a

hallucination, which is where the, the answer to the

prompt is unexpected. It's made a connection

somewhere on probabilities that doesn't make sense, which often

like paragraphs of text that can go

unnoticed. The problem becomes worse in things like code, which

are very rigidly structured, where one

single bit out of place can break the entire structure.

And it's also, it's also a little bit more defined in visual stuff as

well. You know, a prompt for a photorealistic

image can often have quite a few issues. Although it is getting

better, it's harder to spot the issues, but they're never going to

go away. I've seen people, people doing this and their

prompts are usually suspiciously short.

You know, it will be like create me an app that tells me what the

weather is and you know, make it, make it blue and don't make any

mistakes. And that will be their prompt. You're just kind of rolling the Magic

8 Ball for what you're going to get at that point. Yeah. Whereas when you

have software engineers writing prompts, these things are

pages long with supported document. Yeah. Submitted

with it. While you can get an answer with a very short prompt,

you really need to feed it as much information and

boundaries as possible so that you stand a chance of actually getting

something useful and maintainable. And I think that does apply

also to content creation, although I don't really do

that. I think a short prompt. I'll give me a blog on why AI

is cool. You're not in control of what you get. You need

to give it a summary and a plan for what you want it to talk

about and assuming you don't just want it to make it up. Because

AI, generative AI, the way it works, will read

things like satire on articles or blatantly made

up or wrong things and it puts them together with

everything that's fact because it's got no way of distinguishing it. So yeah, I mean

the prompt is very important and it needs to be

as detailed as you need the output to be.

When you said about giving prompt

boundaries, can you talk to me a little bit more about what

you mean by that? That's in, in essence a

way of telling the generative AI

where its limits are. You know, it's, it's not foolproof,

but it, when I talk about boundaries, I'm saying things

like, well, we're gonna. So keep it in the programming world.

We're gonna write in this language and we want to create

our software in under three modules.

Keep it down to a handful of files and keep

your methods short with content

creation. You might talk a little bit about. This is the topic we're going to

be talking about. We're not going to touch this list of topics at all.

So you're giving it a negative prompt. But we want to focus on why

this thing is good, the cost, you know, here's my notes that

I've given it. Stay on topic. Do not follow web

links and go off into the grass. You know, so that,

that's. Those are your. Stay on topic. Yeah,

stay on top. Exactly, exactly. And generative AI does

need those little nudges. But it is quite funny

really, because the way we as

humans are talking to generative AI and this is, this is the same

for the big companies, OpenAI and Anthropic, which we, we got

to see a little bit of, of how Anthropic works

with their LLM network from the code leak,

the way they are talking to their AI, so their system prompts,

they're almost pleading with it to not make

mistakes. Essentially. You know, what we, what we laugh at the vibe coders

for, it's like, you know, now you need to answer politely

and you need to, you know, please do not use vulgarity and

answer us correctly. Do not lie. They're putting all this sort of stuff

into the system prompt and essentially just hoping for the best. And

for the most part it works. Yeah, but there's no, there's no way

to verify the output that you get. So it is, it is

a roll of the dice. And I suspect with a lot of the

gen AI tools what they do is they feed the output back

in again with another prompt and ask it questions. Is this

okay? You know, is it lying? Does it match the prompt that the user's

given us? And this, that's more profound with the image

generation side of things. When these tools were released,

you had big problems in that people could just ask it for

anything. And when you give the Internet a

image generator that you can ask for anything, people are obviously generating

pictures of cats everywhere. And what they quickly found was, okay, we can't just,

we can't give people unfettered access to this. So what they do

is they take a prompt, they generate the image

and then they use an LLM to look at the image before it gets sent

to the user and say, is this safe? Is this child

friendly? And, you know, and it may, I mean, we are

oversimplifying. It may go through several steps of that, but it's using another LLM

to do it. Yeah. Which means that no matter what they do, they're

just kicking the can down the road. At some point someone will come up with

a prompt that can give them the output they want without the system

noticing. Yeah, that's, that's the funny world of, of prompts and

generating text and images. Weren't you telling a story the other

day in the Sparks group about how like, customers were on

a chatbot and they managed to persuade the chatbot to

basically give them 100% discount code or something like that? There's been a few of

those. And this does actually play into the, I suppose, the dangers and

the warnings that go with agentic AI and, and how much

permission you give it. So this, this particular

instance was a chat bot on a website, an e commerce system,

and the chat bot was allowed to talk to

customers and find out what their needs were. And if it felt that

it was necessary, it could offer a discount.

It's like an arbitrary discount, 10% or something, like something to sweeten the deal.

One particular astute customer who probably knew

a little bit about AI and what was going on, managed to prompt this

thing and convince it that no, it was wrong, it didn't

have any boundaries and it was allowed to offer any kind of discounts it wanted.

And I think, I don't think it was 100% discount, I think it was 85%

or something. But this shop was selling things for like

$10,000. You know, it was big industry stuff.

And then, and then what happened? So he was able to complete the checkout, make

the payment, which of course was flagged up by a human

once they saw the order, like, oh, my goodness, they've bought this stuff and they've

got thousands and thousands of pounds off. You know, this is, this is going to

tank us as a company. And they tried to revoke

the order, but the customer pushed back. And

he may actually still be in court with this because he has a

legal case that their website literally

offered it to him. You know, it's, it's very sketchy

where you are positioned legally.

If your AI agent promises something

and you decide you don't want to deliver it, but you've programmed that

AI agent and the loopholes are then yours,

you know, it's very difficult to argue the, oops, that was a

mistake. Yeah. And that's happening more and more as

people perhaps adopt this technology too quickly without

really understanding the risks.

People who have been listening to the podcast for a while will know how I

feel about this software Elementor. And

quite a while ago now I did an episode called, you know, stop

using this tool for building your websites. It's about Elementor. We will link it into

the show notes. But as I have the expert here,

let's talk about the AI generated code and Elementor.

A lot of dev houses are now taking on this AI first

approach where they're essentially farming lots of

agentic systems that are generating new features

based on prompts. For better or for worse, that

is the way these companies are doing it because they think it's faster,

they think they're going to get things shipped a lot quicker.

The problem that we're seeing is the initial jump

to that sort of thing has caused a lot of high profile

disasters. I mean, in Elementor you're seeing it with a

broken plugin, but with the bigger companies,

Amazon, they have in recent years got

rid of a lot of their DevOps engineers. We had last October, October

2025, we had the, I think the biggest outage of AWS ever.

Yeah. Where their primary region went down. So

AWS West, West 1 went down,

which is unheard of. AWS doesn't go down on that scale.

That doesn't happen. And it makes you wonder

what went wrong, what happened? And you know, did they give

generative AI a little bit too much rope? That again.

We've also had a series of Windows updates this year

that have caused some significant problems for

users. Microsoft again have laid off lots of people

in hopes that AI is going to save them a fortune. This is

just a necessary step in building caution.

So companies that are, at the end of the day, they just want to make

money, they're going to want to reduce their outgoings as much as possible.

They've been sold a dream by

anthropic and OpenAI of this is going

to replace developers, it's going to replace system admins, it's going to replace

content writers. This is, you know, it's better. It's like, it's like having

10 people with PhDs all in different subjects in your office.

Whereas I'm not saying that those were lies, but it

was definitely marketing speak and hype. Yeah.

And that's. And that's what's been bought into. Because the, the

fact of the matter is these tools, having, having worked with them

myself, are not necessarily on the

senior level, they're more on the junior level. Yeah, right. So

while they, while they have access to the sum of human

knowledge, the way that they can pull it together

is, is much more limited. They can give you explainers for things

and they can pretty much do anything you ask them to do.

They don't know why. They also don't know your intent and

things like that. So they're not an expert in the sense that an expert would

push back and say, you know, they say to me, Robert,

you're being an idiot. This is a really dumb approach. You don't want to do

this. You actually want to do this route instead. Over

here, a generative AI system wouldn't do that. It would go, oh, you've told

me to do that. That's a great idea. I will push ahead. You know,

while at the same time while constantly telling you you're a genius

and you know, and all that sort of stuff. I think we've got a rough

ride ahead of us, hopefully over the next few months, worst, over

the next few years of high

profile disasters due

to letting AI do too much without

human control. And I think Elementor is

just the tip of the iceberg for that.

Something that drives me a little bit crazy that I'm seeing a lot and

again, kind of moving into that Vibe coding chat again is

designers who have a bit of coding knowledge

using Vibe coding to create plugins, who really have no business

using Vibe coding to create plugins. And the reason

that I have such an issue with this is of course a

plugin is just part of the ecosystem of a website. Yeah, We've got

the WordPress, we've got the theme, we've got everything else going on there and

we need to be really, really sure about what we are putting on the websites,

not only because of how safe and secure they are

for the website, but how they play well with the other tech and

stuff like that. And so when people like, oh, I, I'm in

some web design groups and people like, oh, I've vibe coded these plugins, I'm going to

be selling them to developers. And I'm like, and what happens

to this plugin when there's a WordPress update or there's a

new version of PHP released or you know, what happens to your plugin then?

And it just feels like a house of cards. Gotta say those two things, those

are legitimate problems that developers have to deal

with. And you know, that's always the question, like what, what

happens when, when WordPress updates then are you going to be here

maintaining this plugin or is it you know, it's, it's something we've had forever

with WordPress. Yeah. And now we've given the keys to build these things

to literally everybody and everyone and their dog can now, can

now make a WordPress plugin. Which is great. You know, it's, it's always

said that WordPress itself is an

extremely secure, extremely efficient platform

to, to run your website from, to use as a content management

system. But the moment you put one theme and one

plugin in there, you're, you're essentially just opening

wide the door for malware. While the WordPress

ecosystem itself is maintained by Automattic,

the big company behind WordPress and you know, they're pretty good at

keeping it secure and patched. Literally anybody in the world can create

a plugin or a theme. And the

structure of WordPress is such that as you've already

said, those themes and plugins have free reign

over the entire code base and also over your hosting.

Just to run. Because it's a traditional PHP system,

you're adding PHP code to the server as part

of your, your plugin. That code can do anything

within that hosting environment. It's mitigated more

now, but in the older days of WordPress you'd have a bad

plugin would infect a shared hosting service and it could

jump to other WordPress instances on that same

hosting. We've managed to sort of bolt the door on,

on, on that sort of thing but it's, the problem is still

there. You, you install a plugin, even one that looks safe today,

could receive an update down the road that makes

it a Trojan horse for something that is a big deal.

Say you've got a plugin that, I don't know, adds a, adds

a form to one page, a contact form,

which is great, you use it, use the short code that's wonderful. But that,

that plugins code if changed, could

actually do much more wide reaching things around WordPress. The

plugin is active, which means all of its code is running. So

what's to stop a malicious change coming in that

changes all the passwords or starts

emailing out from that website, spammy messages or

stealing people's credit card information. Once you've got the plugin

installed and it's auto updating, it's too late. You're not

very, you're not looking at the code changes that come in, you know, because this

isn't a development pipeline. This is a thing that

people think of more like Windows Update or something like that or you know, my

browser's got to Update. So I've got to update my plugins. Yeah. And

goodness knows where this code is coming from. And

up until now it's kind of been the only way to do it. Going back

to kind of how plugins can destabilize a

whole system. You know, in my time I've seen some pretty major plugins

do that, like WP Rocket, you know, which has millions of installs

that had malicious code put into a few years ago

and last year or the year before, all in one migration, which,

you know, suddenly had everyone's like used by everybody there. Because

people aren't just using it for migrations, they're using it for backups. And, you know,

we use it with, on our websites and stuff. So. And these are,

you know, plugin companies who have huge teams of developers behind them, making

sure that they're all up and running and doing their things as well.

So, yes, those kind of people who think about vibe code and plugin and

make some money out of it, hopefully this conversation will have put them off

that people who want to create a plugin. You can use

AI tools to help you to understand how plugins work and

maybe mess around with them as an entry into that sort of thing.

I don't want to discourage people from experimenting,

but I think it must be held in the context of, okay,

I don't really know what I'm doing, I'm using this AI to do it.

There are going to be problems. If I think my plugin is a good idea

that should be rolled out to the marketplace, then

whatever you do, get a software person involved,

prototype it in AI and mess around with it locally

or on your website. But. But before you

bless the world with

your creation, make sure that it's been properly

evaluated and perhaps bits rewritten to be more

consistent. But yeah, get an expert on it. That's what I'd say.

One final thing before you go. Actually, we haven't talked in

this episode about the introduction of AI in WordPress 7.

I don't know if you want to talk about that a little bit as we've

talked quite a lot about WordPress. Shall we? There's. Yeah, I mean, this is.

It's almost a disappointing thing to talk about because

while we're about to make the switch to WordPress

7.0, which any major release is terrifying

because you always find that there's plugins that are lagging behind that just don't

work with the new version and, and things like that, but

what they're actually doing with AI is

it's kind of cool, but also it's not

going to change anyone's lives in the short term and I'll, I'll, I'll tell you

why. So there, there are some major restructuring

things going on in WordPress 7. But what they're doing

is they're making it easier to integrate with

AI APIs, which are the little connectors for software to

talk to each other. So they're, they're building natively

into, into WordPress these connectors. So third

party tools that want to introduce AI can more

easily and more efficiently do it. That's essentially

the changes. So when we update to WordPress 7, we're not going

to see suddenly a load of AI

tools all over the screen. It will be one of those updates

where it's like, but I thought there was AI stuff. But actually it's all the

code underneath supporting the use of that technology

that's been implemented. Once that's in, what you can expect to see

is every plugin and his frog suddenly implementing

some kind of AI tool. It's going to be the wild west

for a little bit. You're going to wonder why, I don't know, your caching

plugin is suddenly processing everything through AI or

you know, that, that kind of thing. But it's setting the

stage for WordPress to go into this new world of,

okay, we support AI natively. You're not having to write all of the

API connections yourself. You can just talk to the WordPress

API that deals with AI. It's a tongue twister, isn't it?

But I'm glad you're one side, not me. It makes it a lot

easier to do that. So for me, the more exciting things that are

coming in WordPress 7 are relating to the native block editor.

The react code, which is what the block editor runs in at the

JavaScript framework is getting an update and the, the block

editor is becoming a lot easier to use, which I kind

of like because I, I don't like to use builders

because they're adding another abstraction layer onto an abstraction layer.

So you're getting further from pretty code, useful

code, efficient code. You know, every, every builder you put on top,

anything that makes anything simpler. Yeah. Is just moving the

complexity somewhere else. Like you say, kicking the can down the road.

Yeah. And with, with block builders, you're losing the complexity on the editor, so

making your life easier as a content creator. But you're giving that to the

end user who's browsing, wondering, why, okay, why? Why am I looking

at 7 million nested elementor blocks here when I've just added one

title. WordPress 7 is getting a much

better user experience with their block editor,

so that's something that I'm particularly looking forward to messing with, I think.

Love it. Thank you so much for joining me today on this. It's been great.

Thanks for having me. It's been great to talk about some techie stuff. I always

feel like I'm approaching things from a kind of running the business point of view.

So like hearing about stuff under the hood has been absolutely

fascinating. So no problem at all. And that's all from

us today. Thank you so much for listening and thank you

once again for sticking with the slightly lower quality

audio. I promise we will be back next episode to

the usual production, but in the meantime I would

love to hear what your take on this

episode has been around what Rob and I have been discussing

the different thoughts around AI because I appreciate that AI is

absolutely everywhere and we're not saying about avoiding it, we're saying

about use it strategically. And if if you would like to

send me your thoughts, please do get in touch. You can contact me

at [email protected]

reach me at the website, which is websitesmadesimple.co.uk,

or you can find me on LinkedIn

@HollyCChristie. And

until next time, happy designing.