From playing with AI to building with it, with Dr Dan Maggs

Many people try AI, enjoy it briefly, then struggle to make it genuinely useful. In this episode, John Bennett talks with Dr Dan Maggs about the shift from experimenting with AI to building practical tools, and what that makes possible for non-technical founders.

They discuss how AI has moved from novelty to something you can actually build with, why context degradation causes long AI chats to break down, and how working with projects and workflows helps address those limits.

Dan shares his own journey, from early experimentation to developing a working meal planning app, despite having no formal coding background. The conversation also looks at choosing AI tools without chasing every new release, using AI as a non-judgemental sounding board, and what this shift means for people who want to build personalised products.

The episode closes with a listener question on structuring AI for complex tasks like business plans, and why thinking in terms of projects matters more than writing ever-longer prompts.

In this episode:

• Why AI often starts as a novelty and then disappoints

• What changes when you add context and think in projects

• Moving from prompts to building real tools

• Building applications without traditional coding skills

• Context degradation, and why AI chats “forget”

• Designing around AI limits with apps and workflows

• A real example, building a meal planning app

• Choosing tools without chasing shiny objects

• AI as a non-judgemental thinking partner

• Listener question, structuring AI for business plans

Transcript

John Bennett (00:19)

Welcome to the AIQUALISER podcast where we try and balance out the AI noise with conversations about how people are really using it. And today I'm talking to Dr. Dan Maggs who was the person that showed me how useful AI could really be. So welcome, Dan.

Dr Dan Maggs (00:36)

Thank you very much. I'm so honoured to be your very first guest on this podcast and I hope it delivers.

John Bennett (00:43)

Well, I think it was natural it had to be you because as say you set me off on the journey. So, there's so much has changed in AI since we last spoke and I'm really excited to get into that but maybe first you can just tell us a little bit about who you are and what you do.

Dr Dan Maggs (01:00)

I think that's a really good question actually because I'm not really sure I know the answer to it because as you so rightly point out, the whole AI space moves at a pace, doesn't it really? And so, you're right that so much has changed in pretty much the year since we were last together and we kind of had that conversation that sparked something in you, which was really cool. But yeah, I think...

even potentially what I've done do in the last three months has probably changed as a result of AI, which is pretty kind of cool really, isn't it? So, I mean, the background is that I'm a doctor, I'm a GP by training. I think I've always known that being a GP wasn't everything I wanted to do with my life. That's not that there's anything wrong with it. It's just because I've always had this kind of entrepreneurial spirit and things like that. So,

From my perspective, back in 2016, 10 years ago, I lost a significant amount of weight and decided to start a YouTube channel, basically teaching everybody what I wanted to do. And that's actually how we met because that's what introduced me to your brother and we met through that really. so coming to basically cut the last 10 years down to something very short, ⁓

grown this YouTube channel, I've had a coaching business sort of off the back of it with varying degrees of success. And then I've started to get very interested in the AI side of things. And, you know, it started very simply in terms of like when ChatGPT launched through to doing slightly more complicated things. And more recently I've essentially become an AI developer really, and started to build out my own

plural apps actually. I've gone on what for me has become the next logical step. I think that's probably what we'll end up talking about.

John Bennett (03:11)

Sounds really interesting. So when did you first get into AI?

Dr Dan Maggs (03:17)

I think I got into it at the beginning. It's just what's been interesting is learning.

what to do with it because, you know, is it that useful at the end of the day? It's very novel and very interesting and quite good fun, but I think my interest is now is how we can make it kind of useful.

John Bennett (03:39)

And that's one thing you showed me because back at the start, I started to use it and I very quickly got bored because I did a few of those fun things and I think I had a mastermind group when I had the Dalai Lama and Mr. T giving me advice on things and really quickly, it just felt shallow and it just felt like I wasn't really...

it was a play thing really rather than something serious. And then when I bumped into you last year and you showed me the work you were doing on personas and business profiles and that's what really opened my eyes and got me to look at it again. So was there a point when that same thing happened for you that it turned from a bit of fun into something more useful?

Dr Dan Maggs (04:23)

Yeah, and I think I've been learning from people along the way who've kind of shown me what can be done. And so I think, you move in from just typing random stuff into a box and seeing what comes out to starting to put a bit more thought into the prompts that you're putting in there in terms of, you know, how you get it, what you want out of it. And, you know, I was trying to write YouTube scripts and

you know, it did help, but it wasn't perfect. It wasn't brilliant. And then you start to kind of play around with, you know, adding in context. So it knows a little bit more about you. And it was this, that was kind of what we were doing last year in terms of setting up projects where the project is there for a specific person, purpose, and the project is therefore has got.

context, you can upload documents and all that kind of stuff really. And then I felt like that was when it started to become genuinely useful in terms of the kind of generative AI side of things really.

John Bennett (05:33)

Yeah, that makes a lot of sense. And you mentioned a minute ago, which I didn't know, which is that you've kind of moved on. I last time we spoke, you were using it in that way. You were using it to help with the YouTube channel. So what's the journey from there? What's happened since?

Dr Dan Maggs (05:49)

So I think I've come from this, I've always had somewhere on my to-do list for as long as I can remember, learn to code. And so now I never got around to it because of one thing or another, but of course now you don't really need to learn to code anymore, I don't think. And so what happened was I was redoing my coaching business towards

I think it was May time last year and I went away on holiday but I was away a few days earlier with my daughter. My daughter is four years old. So she was in bed early and so there was me kind of sitting up going, hang on a second, I just got some time here. So I got my phone out and started talking to Claude, I think it was, and Claude, the question was, well, so I've got my

my website and I want to turn it into an app. How do I actually do that? And it was like, well, you're going to struggle a little bit because actually your website is based on WordPress. okay, fine. So WordPress has basically been the kind of default standard for building a website for most people, normal people for the longest time now really. And I was like, well,

okay well if I'm not going to use wordpress then what am I going to use? It was like okay well yeah I'd probably recommend something like a more modern web stack like Next.js and with the super base backend hosted on Versal and I was like well yeah but I don't know any of that stuff and if anybody listening is going what the hell is all that kind of stuff well that was exactly my reaction at that time. Well yeah okay that's this is basically this type script kind of thing and this is what most of the modern

web apps are all built on and stuff. I'm like, well, yeah, but I haven't got a clue how to do that. It was like, well, don't worry, I do. And so all of a sudden that kind of possibility opened up to me. And I think whenever I was using AI, I was always very aware that there were other people using it in a way that was more advanced than what I was using here.

And so I was always very aware that you can be using AI in a kind of coding context, but it was just like, it was a complete brick wall to me in terms of I hadn't got a clue what that even meant really. And so I think that is kind of where I feel the next way is going with AI. And I think where I have the issue with like,

I feel like I've come as far as I can go with writing ⁓ prompts, giving them context and all that kind of stuff and making them do useful things. In order to move on to the next stage, there's got to be something else out there really. I basically sat in the evenings during a holiday while everybody else was asleep, rebuilding my website and then I...

John Bennett (08:58)

Hmm.

Dr Dan Maggs (09:07)

started building my wife's website out to replace hers and that was quite useful because my wife's a singer and so she got this diary on a spreadsheet and it was like well she was paying her mum to do the admin work to start to copy the things from her diary into the website for her kind of tour dates.

There was mistakes being made and whenever a gig got canceled, it would be an email to say, oh can you please take that one out of the diary and all that kind of stuff. And it was extremely faffy. And so I built out using AI and code this link between the two. And so now it's a case of if Vicky wants a gig putting on her website, all she has to do is, well, she doesn't do anything because it's automatically.

in her own backend structure. And was just like, well, okay, well, if I can do that, what else can I do? And so the key kind of thing came from me was realizing that I don't need to learn to code anymore. And then I'm very aware that, you know, a lot of the stuff we're talking about here is quite timely, but around that May time last year, Anthropic, who are the people who make Claude,

Probably the, in my probably fairly ignorant knowledge, the probably main competitor to OpenAI came out with, who made ChatGPT, came out with something called Claude Code. And so Claude Code is basically this, it basically allows you to just talk to it and it lives in the software sort of thing and it just writes the code for you. And so I just started to kind of play around with that really.

John Bennett (10:46)

Mm-hmm.

Dr Dan Maggs (11:02)

And I think what's become very interesting is some of the things that I was struggling to make work in the kind Claude desktop environment or the Claude web app, all of a sudden became possible in terms of you're actually doing them within code. Because I don't know if you've noticed, if you work on a, ⁓ say you're working on a document.

within a chat, within Claude or OpenAI or ChatGPT or any of your favorite kind of apps, they start to, after a while, get a little bit more stupid. And they start to become, and they start to kind of forget what you've told them earlier on in the prompt and the chat degrades, if you like. And so I call it like a digital dementia.

John Bennett (11:58)

Hmm.

Dr Dan Maggs (12:01)

But I think you could also think of it as it becoming tired, if you like. But actually what you're seeing is called context degradation. And that is that actually the more kind of context it has, the worse its outcome becomes.

John Bennett (12:02)

Right.

Hmm, yeah.

Dr Dan Maggs (12:24)

So I'll tell you what I'm currently working on at the moment, hopefully launching within the next two or three weeks. So there's a meal planning app, okay? Meal planning app for families.

John Bennett (12:32)

Mm-hmm.

Dr Dan Maggs (12:33)

So one of my big things is obviously I'm into their kind of health space and the nutrition space and the weight loss space But actually one of the big things I just wanted to do is make meal planning easier for my family. Okay, I Follow a pretty low carb diet. My wife doesn't and my daughter follows an entirely different kind of situation And that's the reality that a lot of people face, isn't it? Like, you know

John Bennett (12:54)

You

Hmm.

Dr Dan Maggs (12:58)

People have

different likes. My wife hates mushrooms. I love mushrooms. My daughter's still not sure about mushrooms. And so, you you're coming up with these meal plans and yeah, I actually want to be able to eat better, more easily. I want to not have to go shopping three times a week because I've forgotten something the first time I went or I didn't plan things out. And so this was something that I was struggling to build out in Claude Desktop because of this context degradation problem that

we've been having. It's doable and I got some lovely meals out of it, but it's not easy and I wanted it to be frictionless and easy. So all of a sudden I'm learning these skills through what I spoke about earlier, which is Claude Code and building my wife's website out. And it's like, well, actually I think I can make this now. I think I've got the knowledge to be able to make it. And so turns out I have and it's

John Bennett (13:29)

Hmm.

Dr Dan Maggs (13:57)

very, very nearly ready, I think within days of being able to get out to people beyond myself really. So I've learned to be a developer of sorts really. And I still feel very much an outsider in that space because I'm not a traditionally trained kind of developer and I'm just somebody who's just sitting on his own computer kind of trying to play with these kinds of things. But I think

John Bennett (14:09)

Hmm

Dr Dan Maggs (14:26)

I heard a quote from Sam Altman, who's the CEO of OpenAI, and that one of the great things about AI nowadays is it's making these kind of technologies that would have only been available to a select few, available to the many. And I think he's right. I think he's right. But I think if we stop our thinking at what can I do

John Bennett (14:44)

Hmm.

Dr Dan Maggs (14:53)

in the desktop version of ChatGPT or Claude or the app that's on your phone or whatever, you're going to limit where you can go with this. Because I think what the next thing is, is how can we solve real problems in a way that we couldn't do before the AI era? Actually, when you start to dig into it, it is huge. So if you think about it, John, what's the previous,

the standing model of business in Silicon Valley. It is have an idea, pitch it to some investors for an initial round of funding, go away and build a group of developers who will actually then build out a very first version of that product. Well, actually you don't need that anymore because if you're a reasonably motivated, slightly intelligent individual who is able to go and

have an idea, you are now able to then go away and relatively inexpensively actually go and get to a working version of a product relatively rapidly. so, well, first of all, that's gonna disrupt the entire model of Silicon Valley. But also, it's also gonna allow

problems that were previously too expensive to be solved to be solved and I think that one of the things that people are gonna have to get their heads around is yes these tools are available to the everyday person however that doesn't mean they're necessarily easy to get your head around or use and I think that's the the key thing and so I was thinking about what

John Bennett (16:27)

Hmm.

Dr Dan Maggs (16:49)

how our conversation might go. And thinking about my own evolution from those first days writing silly poems in the style of Donald Trump through to what we did last year in terms of helping you build out a ⁓ website for the business that you were working on at that particular time through to actually where that line finishes in terms of what is useful

John Bennett (17:11)

Hmm.

Dr Dan Maggs (17:18)

in terms of what you can do in an app, in one of the desktop apps versus where it might be able to go in the future. And I realised that this stuff is going to date very quickly, but I think the key thing is now that we are able to give, we are able to have these AI tools interact with other AI tools or interact with any other things. We can have them send emails, check a weather forecast.

or check a database, for example, or write to a database and actually do genuinely interesting pieces of work, semi or completely autonomously. And I think that's where we've got to get our heads around. And because currently the only people who really understand this are people who are technical people, people who are from the developing, the developer.

John Bennett (18:02)

Hmm.

Dr Dan Maggs (18:16)

side of things, people who've learned to code and in the traditional sense. And actually most people I think are using AI in ways that is way below what its capabilities are. And so I think the real interesting thing is going to be when non-technical people go and start to learn this kind of stuff. People who are working in marketing, people who are working in healthcare, people who are working in all sorts of different areas.

John Bennett (18:30)

Hmm.

Dr Dan Maggs (18:45)

who are then going to be able to go and learn this kind of stuff and go, hang on a second. I know that there's a way to solve a problem here using AI, because the people who know about how to solve those problems don't know that those problems exist. And so I think we're at an extremely exciting time in history whereby there's going to be a meeting of this new technology with people who are able to actually fundamentally use it and

John Bennett (19:02)

Hmm.

Dr Dan Maggs (19:14)

I think that's going to be a massive growth space in the next kind of, well, I would say year or two because the timelines are extremely short here, really, aren't they? With AI, everything moves at such a pace. But yeah, it's very exciting when you start to be able to see what is possible when the other things are available, really. So yeah. Anyway, that was a long answer, wasn't it?

John Bennett (19:25)

Hmm.

Hmm. Yeah.

It's great. And there's a few things there that I really wanted to pick up on from. And there's a whole range of different nuances, if you like. just the one simple one first, you talked about the context degradation, the context window, and when it starts to get a little bit forgetful and get things wrong. And when you're working on a document, in some ways it's easy to see that because

it's quite easy to look at the document or the response and go, no, that just doesn't sound right. So when you work it on code, that must be much harder to spot, I guess, when it starts to be forgetful. How do you keep an eye out for that?

Dr Dan Maggs (20:20)

I think what it is is actually that what, well certainly from my experience so far, and that is limited, honestly, I'm new to this, don't complain, I don't claim to be an expert in this, and I'm very aware of that unknown, unknown thing where I don't know what I don't know, and therefore, you know, people who are kind of experienced developers may turn around and go, no, that's stupid, but.

The way I'm kind of building what I'm building at the moment is so that context degradation doesn't necessarily happen at all in terms of, you know, so for example, linking in with what I've said, okay, so previously I would have given Claude, I would have set up a Claude project and this is what I did initially. And I would have up a Claude project and I would have given it

several documents. I would have put a document in there which said, these are what our family likes to eat. Okay. This is, you know, my wife hates mushrooms. I like mushrooms. My daughter's not sure about mushrooms. There are three of us in our family, you know, are and we're all these age, et cetera, et cetera. And so give it all that context, um, behind there and then give it a set of instructions, which is like,

Well, initially, I want you to go and ask me what our schedule is for this week. Once you've got the schedule, what is the... Have you got any guests for the week? After that, have you got anything in your fridge that needs using up this week? And after that, have you got anything in your freezer?

you were potentially giving it a list of stuff that was in your freezer that were in the document, which are the stuff out the freezer. Have you got any meals that you want to try from your list of meals that you want to try? Yeah, okay, I want to try these. And then all of a sudden you've got this long linking of things where it's already starting to break down even before you've given it the, okay, go off and write a meal plan for me. And before you've gone to say, okay, right, now generate those recipes for me.

And so you've already got a situation there where it's losing degradation. Compare that to the situation where all that data is in a database. Okay? And what it says is, I've, what's your schedule? It knows our typical family kind of school term time schedule. Knows what our schedule is. Okay, here's your schedule because it's then pulled it out of a database instead of

somewhere else. And then, are there any deviations from that schedule? No. Okay. Move on. And then it saves that schedule back to the database to say, and so you move through those steps and instead of it just growing the context window, it's then saving it back to the database. So when you come to say, okay, we'll generate my meal plan, what you're then doing is actually building a new prompt in a completely fresh context window.

which then pulls all that data that you've given in the first few stages out into a ⁓ new fresh prompt. So when it generates that prompt, it's generating it from fresh information rather than what you wrote 10 chats ago. that's the problem there is the fact that it's forgotten that you wrote the

I'm working on Wednesday evening and so I'm going to eat on the fly while I'm out rather than at home. Well, it's forgotten that, but actually instead it's actually just and it's just like you've just built one big long prompt essentially at the beginning. But it's that and so fundamentally it's not that I'm watching out for context degradation. It's that I am eliminating it by design.

John Bennett (24:33)

Hmm. It makes a lot of sense and and presumably that's something you can apply when you're working on documents rather than code as well.

Dr Dan Maggs (24:33)

that makes sense.

I think so, yeah, I think so. Typically, when I was writing YouTube scripts with it, I was actually breaking it down and I was actually having multiple different projects with different elements of the script writing process. And so it would be a case of, okay, well, we're trying to come up with the idea for a script, all right? And then you would have a document where it gave you that and then...

you would then take that one and post it into the next one, which would then have a fresh context window and would do a different job. And I think that that's a way to kind of semi do it, but it's not a particularly elegant way of doing it in terms of scalability.

John Bennett (25:28)

Yeah, it's something that I was interested in to talk to you about there because when we last spoke, we were both using Claude. I do most things in Chat now. And the reason I do that is because you can use the structure of projects and also custom GPTs. So generally, I'll put those documents that I've got. I might have, for example, a marketing project or ⁓ a project for this podcast. And I'll put the...

the information, the documents in there. And then I've got a series of custom GPTs that I've made over time. So you know it might be one to try and work out the hook from a piece of information or to work on a script or that sort of thing. And as I say, that for me, Chat's working great like that now, but I've not looked at Claude again for a long time and I know they've got Claude skills now. And I just wondered what's your experience there.

Dr Dan Maggs (26:16)

I think in a way I've stopped looking. I cancelled my ChatGPT account, my paid account, eight months ago now because I found myself living in Claude all the time. I think, yeah, gosh, there's features coming out all the time, isn't there? But it's like what is actually helpful for you?

And so like, ultimately, like, there's a lot of good tools out there, isn't there? Like, I occasionally jump into Perplexity when it's research based stuff. What's the Google tool that is it Google Notebook LM or something I think it's called? Notebook LM, which is absolutely fantastic when I'm doing a research phase of script writing, because I can literally drop in 30 or 40 documents.

John Bennett (27:02)

Notebook LM,

Dr Dan Maggs (27:13)

complicated scientific documents and query it. And it's phenomenal for that kind of like research thing. like different tools for different things, isn't there? And so like, and some of those, for some of those things, the native apps of those LLMs will be brilliant as they add features to them. But it just depends on what you're, you actually want to be able to do.

John Bennett (27:26)

Hmm.

Dr Dan Maggs (27:43)

but also where the limits come in. I think, you know, if you talk about the kind of one of the things that came out, I say relatively recently, I don't know how relatively recently it is now, it's probably within the last year or year and a half, is the deep research functions of ChatGPT or Claude or things like that. And what that is is a great example of an AI agent, okay?

And so an AI agent is something that can essentially got a set of tools that can choose its own pathway to come up with a result. And so when you type in one of these deep research prompts, actually there's multiple different ways that it can be solved. And so it has got the ability to go away and decide what it needs to do. Plan a course of action.

and launch all these different tools in order to actually then get you some sort of aggregate response in terms of a decent output at the end. And I think that's the buzzword in kind of AI over the last year has been like these agents as well in terms of things that can autonomously do a job. You don't necessarily get that

in your, other than the deep research example, within these native apps. But you can, if you've got a specific use case for them, build out those kind of things. And the other thing is workflows, is you've got different, this needs to happen, then this needs to happen, then this needs to happen. Like what I was describing with the AI, the AI meal planning stuff. First of all, decide, get the user schedule,

Find out if there are any deviations and save it back to the database. Next step, you know, and so on and so forth. That isn't as easy to do in these kind of native apps, but it's dependent on the workflow, isn't it? And I think that's, it's always good, I think, to be able to learn about what other people are doing in your industry or in your world.

And also then go and look over the shoulder of people who are working in completely different industries and go, hang on a second, is there something interesting and exciting that actually applies to my world here? Because then you start to kind of get that cross-linking kind of stuff with regards to what's actually possible. I do think we're in a phase where it's like, what is possible here and what is valuable? And obviously with the example you've just given about the

John Bennett (30:13)

Hmm.

Dr Dan Maggs (30:38)

the custom GPTs, that is valuable for that particular use case, but not necessarily. And so in answer to your original question, it's like, well, I think it's easy to be going, which is better, ChatGPT or Claude? Hmm, well, depends, doesn't it? You know, or Perplexity and people have got their favourites, haven't they? But it doesn't really matter.

John Bennett (30:41)

Hmm.

Hmm. And that, yeah.

Yeah, and I think there's a danger isn't there of chasing the new tools all the time? You know, as a I say I've heard of Claude skills, they may be able to replicate the things I'm doing in Chat, but I'm quite happy where I am right now. So there's that tension of, is this something genuinely new or is this just another distraction? How do you approach that?

Dr Dan Maggs (31:11)

Mmm.

Mmm.

Yeah.

I mean it's interesting isn't it with an open mind of course because you're telling me about Claude skills and it's like well actually I don't know what that is so okay so what do we do we open up a new window and we have a little look at what Claude skills are do you know what mean and so I think we have to kind of be careful about kind of jumping down rabbit holes and stuff but I guess you know unless you jump down a few rabbit holes you guess you don't learn do you really so it's like

what's the unknown unknown? How do Claude skills? How do I know if Claude skills are going to be helpful for me or not without actually knowing what it does really? so yeah, we have to, I think you do have to have that open mind, but also you have to be very, very careful about, we're in a, we're in a huge, everything's happening so fast, isn't it? Like you have to be extremely careful about.

just switching all the time because ultimately you just end up chasing the latest thing and actually not getting any work done. And so, you know, if you find something that's working for you, great, but also keep an eye out to the, to the next kind of thing really. So yeah, absolutely. And at the same time, there's loads of new apps launching all over the place, which are promising the world. mean, the, the kind of word of 2025 was vibe,

vibe coding, wasn't it? Which I think is, I don't like the term, but I guess it's what I'm doing, if you strictly kind of take it in that line. But then there's loads of apps, isn't there, out there that are supposed to help you build your app. You just describe your app and we'll build it for you with, yeah, I don't know. Didn't work out that way when I tried it, so.

John Bennett (33:15)

Yeah,

I must admit I've tried Lovable some time ago and another one I can't think of its name and I didn't, I could get some basic stuff done but I must admit I struggled to do things more iterative so I'm really interested in looking at what you've done with Claude Code there when you finished it and giving that a play.

Dr Dan Maggs (33:20)

Mm.

Mm.

Yeah.

of course, talking about how things develop, like, Claude Code comes out with their thing, but also then OpenAI have done theirs, and Gemini and Google have done theirs. It's very interesting, isn't it? Because they all do it, and they all do these things. And it's like, actually, they're comparable tools. And you can get into the nitty gritty of

do you know what? Is the Gemini version better than the Claude version? Or is the OpenAI Codex version better than the Claude code thing? Well, if you're kind of happy with your tool and you're getting stuff done, then do you even need to go looking? So it's kind of having one eye open to new things, but also being cautious of shiny new object syndrome, isn't it, I think?

John Bennett (34:30)

Yeah, that's really good advice. The other thing I wanted to pick up on what you said earlier as well, because you were talking about how we're at this junction now where people that know what the problem is and what needs to be solved are getting the tools to solve it. And what's great in what you've said is that you've taken, for example, the meal planning, you didn't know how to code, and you've gone out there and you've worked out how to do it.

And I just wondered what advice you had for people in the early stage of where you were. They know there's a problem, they want to develop something to work on it or they want to do something new and they're a bit reluctant of getting started. What advice would you give people to how to do it?

Dr Dan Maggs (35:17)

I think one of the things that I think I love about AI the most is it's a completely non-judgmental sounding board. And you are able to go and have the stupid conversations that you might not be willing to have with an actual real person with an AI agent really, or just an AI chat. And so...

It's been brilliant really because all of the stupid questions I've been afraid to ask over the years, I don't need to be afraid to ask them anymore. so like, yeah, you might have been unwilling to be able to get started with those kind of things. But if you want to know how to start an app, you just go into whatever AI it is and say, I've got this idea, what are my options for starting an app? Great, okay, fine. Okay, well.

Now I know the options, but how do I do it without breaking anything stupid? Doing, okay, well then actually that's the next question to ask, isn't it? Because then you learn the workflows and the work processes that people or developers have been using for like years now to stop stupid things happening. And you know, like it's always being able to look forward and

learn about what the next step is. And so if you're naturally curious enough to be able to ask those questions, I think you will get the answers you want. Hopefully it's going to be the right answers. But yeah, guess I just, there's that side of things and then it's just clicking around. you know, looking at when you're in the backend of these kind of like tools and stuff like that, there's that, well.

actually what does that setting do? And I'm okay, I don't know what that is. So I just screenshot it and post it into Claude or whatever. And so what does this do? And so, you know, we're constantly on this kind of learning loop and yeah, there's been a lot of learning for, in terms of the like the app development side of things to where I've got at the moment, but.

John Bennett (37:16)

Hmm.

Hmm.

Dr Dan Maggs (37:33)

actually that speed of learning is massively increased, isn't it, really, these days? And the loop of feedback loop is so stupid. I've got three screens here. One big one on the left, that's for documents, one my main window, and then my laptop over to the side. I've generally got an instance of Claude open where I'm just talking to it to say,

John Bennett (37:41)

Hmm.

Yeah

Mm-hmm.

Dr Dan Maggs (38:03)

that what

I'm doing and learning and actually you might have another instance or two instances of Claude code running on the different window and then like whatever it is I'm building or that over on the other window and switching between them really but yeah I think we're yeah it's very exciting in terms of just having that that you feel like they're people don't you it's clever Claude

John Bennett (38:16)

Hmm.

Dr Dan Maggs (38:31)

being named like a person's name because you feel like it's actually a person that you're having a conversation with, but this like non-judgmental kind of sounding board and stuff. So yeah, like the tools are there. If you've got an idea, you've got a sounding board to go through it with.

John Bennett (38:32)

you

the meal planning app is almost finished. What's next?

Dr Dan Maggs (39:00)

Mmm.

So I've got a few different ideas and these are what I'm doing at the moment is scratching itches that I've wanted. so one of them is like script writing for YouTube and stuff. The frustrations of this context degradation. I would like to be able to build that one out. Then there's something that not necessarily.

probably for commercial reasons, going to talk about. But there are it's basically if you've got an itch to scratch in terms of you're frustrated about how something is working and you potentially got this idea of how it might be able to work because of what you've learned about. Well, then you've got a meeting of something there. And I think

That's what I'm most excited about and where we're at at the moment is because, you know, like just a meal planning app for starters is just, you know, there are meal planning apps out there. I just never found myself using them because they were tedious. But actually there's a new way of building these things from the ground up. And actually it's absolutely doable to be able to go away and build your own app. And actually we talked about that example earlier of Silicon Valley.

you know, a non-technical developer, a non-technical founder would have to go away and work with a team. It's an incredibly slow process, as you know, you know, because you've got to then go, you've got the original stakeholder who's then got to go and have these, you know, the developers, you've got to communicate that to your developers. They've got to come up with an idea and they've got to go away and then bring that back. And then...

You know, it could take months to be able to get anything up and running. And I started this meal planning app three weeks ago, and we've had Christmas in the meantime. You know, admittedly, I did be a bit of work of it over Christmas. But there's something extremely cool about the idea that, you can just you can just be working on it then and there, and especially if you're customer number one and you're building it for yourself. Well, then you're just fixing the things that don't work for you.

you're not going through this idea of, hang on a second, I've got to go and ask a developer and I've got to communicate that problem to a developer who's then got to go away and think about, am I actually going to do this? Do it, then do it wrong and then come back to you. And so I feel like the work that can be done by a non-technical founder in 2025, 2026 can actually be equivalent to a

an entire team. I might be extremely, I might have people wagging their fingers, you know development teams and stuff like that, you know, at me here. But, but actually I feel like there is something quite magical about the way that it gives you instant feedback in terms of you've literally working on the thing that you want to make work and you're building it for yourself first. And that's, that's pretty cool really. And

John Bennett (42:13)

Hmm.

Yeah.

Dr Dan Maggs (42:25)

⁓ and relatively, I'd say not inexpensively compared to the salary of a team of developers. When you think about even if you're on the top tier, the 200 pound a month tier of Claude to be able to build this stuff out fast, it's pennies in comparison to what it would have been many years, even two years ago.

John Bennett (42:26)

Hmm.

Hmm.

I think there's something really interesting that you said there as well, which is that, you know, it's not just about building a product, is it? It's also, you could build a product just for you. Yeah, something hyper-personalised because, you know, there's no way you would get a developer to build a meal planner for you and your family, but you can build one that actually just works for you, which, you just wasn't possible before.

Dr Dan Maggs (43:15)

No, no, absolutely. And I think that is also one of the potential upsides of what AI can do is it can give us the tools to solve problems that were not economically viable to solve before. And so, you know, this is why I think, you know, for me, it's anybody working at any kind of role within ⁓ any sort of company, so basically anyone.

should potentially be thinking about learning this stuff because all of a sudden you get to solve problems that you maybe you didn't know that you didn't even know that it was a problem. It only became a problem when you actually started to connect a few neurons in your brain to go, well, actually, maybe there's a better way of doing this. It's just these are the established systems of how we've done this. And actually, there are frustrations in that

system that I don't like but actually if you don't know that there's a way of solving that well you don't know do you and also that can be done at a very small level and you know very industry specific you know and there's very ⁓ very industry specific stuff stuff that would only really be quite esoteric to us you know to a single company perhaps you know

even a relatively small company, you know, in terms of what they are able to be able to do that would then be able to give them a massive competitive advantage as well. And I think you're going to see a lot of kind of like natural selection over the few next few years, which is where the companies that are embracing this technology are going to obliterate companies that don't.

embrace this technology. I mean, just wipe them out overnight almost because all of a sudden they're going to have such a competitive advantage over the other companies that, you know, and that, that, you know, you just can't ignore this stuff really. And so I just think that the world's changed a lot in the last two or three years. And, you know, my sister's currently looking for a job. She's in a communications kind of

industry and you know she's I offered to talk to her about the same stuff I talked to you about this time last year which she doesn't seem that interested in and it's like well actually your industry's changed massively you know you've got to be embracing this kind of stuff you've got to think more deeply about it you've got to be willing to go and learn this stuff if you want to stay competitive it's it's a very interesting time but curiosity is everything isn't

John Bennett (46:08)

Hmm, yeah, absolutely. Absolutely. I think it's great advice because I think there's sort of in my mind, there's three approaches you can take with AI. You know, can go, you can bury your head in the sand and hope it goes away. You can embrace it wholeheartedly and just, you know, just take what it's saying without question. Or you can do, think, which is what you're advocating, which is to be curious, to learn it, to understand its shortcomings and work out where you can apply it and...

that to me seems to be maybe what we should all be doing.

Dr Dan Maggs (46:43)

Yeah, I mean, it's different things. I was having this conversation with my wife last night. She said, it's not for everybody, isn't it? I'm like, no, it's not for everybody. My wife's a singer. She's incredibly talented at it. Does she need to be learning this AI stuff? As it stands, I don't think she does. But actually, it depends on your industry. It depends on what your role is and it depends on what your goals are, doesn't it? And so I think it probably applies to far more people

than actually are currently embracing it. The bury your head in the sand thing, well, do know what? I think, fair play if that's what you want to do. Just taking it in a very spoon-fed way just might be exactly what most people need. I mean not everybody is going to break ground. Not everybody's going to be an innovator. But also, if you're not, you also run the risk of being

obliterated by AI I think as well maybe the safest way is to kind of be the curious but I think if you're listening to a podcast about AI that pretty much selects you to be one of those doesn't it really

John Bennett (47:59)

Yeah, that makes a lot of sense. That makes a lot of sense.

Dr Dan Maggs (48:02)

Especially if you've got this far into the podcast.

John Bennett (48:05)

Yeah, absolutely. Well, hopefully people are still with us.

So one of things I'm hoping to with the podcast is to answer a question every week if we can that people have sent in. And I've got one here if you don't mind if we tackle that. This one's from Ben. And Ben's question is, what tips do you have when creating and structuring a complex prompt for a project? For example, an in-depth business plan.

Dr Dan Maggs (48:27)

for it.

Okay, so there's several ways to go about this. So thinking about it, I think is only the prompt is too small. I think you probably need to be thinking about the wider context that is available to it in terms of I would genuinely be thinking about some sort of project environment where it has supporting documentation, where you're able to have potentially

you know, for example, a customer avatar, okay? And your kind of maybe your mission statement for your business and all those kinds of things, which the AI is then able to actually draw on to give you more context. And I think this is really what we talked about this time last year and it was having those documents. But in terms of, okay, and so say, for example, like a tone of voice document for how you want it to sound, okay? Well,

actually, but then how do you then go and get those documents in the first place? And so you'd probably need some sort of structured prompt in order to do it. But honestly, I think there are many there are multiple frameworks you can use for that. But the easiest way is to ask AI, the AI you're using. What makes a good prompt for this? Help me write the prompt that's going to get the outcome I want.

is the is the best advice, I think, in order to do that. so getting the AI to write the prompt that ultimately you give back to the AI sounds a little bit circular, but is also extremely helpful. I think, you know, if you can give it examples of what good output is, And I think the other thing is like conciseness is important as well.

Because we do have a very real problem with like context degradation and trying to get a very specific outcome is very helpful. And so being able to give examples of what that specific outcome is, is also really helpful. So, yeah, so context, asking the AI, giving specifics and allowing it to be very focused is good.

John Bennett (50:57)

Yeah, and there's...

Yeah. So I think, are you saying it's about the process really as much as or even more than the prompt?

Dr Dan Maggs (51:15)

I think if you're just doing a prompt, you're missing a big trick. I think you're missing a big trick and the big trick is giving something the wider business context, especially in that example. If the AI doesn't talk about structuring a prompt for a business plan, if it doesn't know what your product or products are, if it doesn't know what your ⁓ ideal customer avatar is,

if it doesn't know what your current marketing strategy is. so kind of the more context you give it within reason, behind the scenes, the more, the better an outcome you're gonna get from it. And I think if you're just gonna give it that in one shot, you're gonna be into a potentially extremely long prompt really, aren't you? So yeah. ⁓

John Bennett (52:10)

Yeah.

Yeah, that's good advice. So I think, and again, this may be where it's different in different tools, but I think probably the way I would approach this is exactly as you say, you need to get those foundational documents, your customer profiles, your offers, your business profile. I'd probably then put that into a project, maybe a business development project. I think for me, the next step would be, when we say business,

plan, who's it for, is this a business plan for trying to raise a loan or try to raise investment, is it a strategic plan? And then maybe try and get some examples of what good looks like for those as well. abstract from your industry, from your business, what is a good business plan for the structure of it, how the depth it goes into for a loan application. And I think I'd probably build that into

a separate tool, like a, I probably have a project for my business information, and then a separate custom GPT that understands that kind of business plan I'm trying to make, and then it's almost like you've got that tension then between the bit that knows your business and the bit that knows what should be in the business plan of the type you're trying to make, and then try and build the two.

Dr Dan Maggs (53:15)

Hmm.

Yeah, I think when...

When we built out last year, what we built out effectively for you is a master marketing hub, I think, from memory. And I think a master marketing hub is great for a business. But I think then what you have to then think about is more specific projects for more specific things. And so the more specific you can be and the tighter your context is for that.

John Bennett (53:33)

Mm-hmm.

Dr Dan Maggs (53:54)

particular thing, the more focused an answer you will get out of it. And so being aware that, yes, there is a time for a general tool, but also being aware that, yes, a specific tool will also be probably more accurate in terms of what you're doing, what you want out of it.

John Bennett (54:19)

And when you say specific tool, are you talking about an AI platform? you talking about a project or GPT or skill within that?

Dr Dan Maggs (54:28)

I mean,

That could be either really. I mean, so, ⁓ you know, for example, like I, with my kind of, my YouTube kind of stuff, I have a general marketing project for the business, but also I would have a specific one for coming up with ideas for YouTube videos. And writing, coming up with an idea for a YouTube video is a specific skill.

And actually then coming up potentially go as granular as this is the idea for the YouTube video, then a specific one for what a title for that idea might come. And the two are distinct different things. The idea for a video and the title for a video are kind of two distinctly different things. And actually one you could potentially give like a bit more free rein to. And then the other one you might just give some

examples of titles or title formulas that have work that work across multiple different niches, but also a separate document for title formulas that might have worked for you in the past, as well as your kind of who your YouTube channel is for. And so I would expect those to do better than also than also the the other things like that, really. The other

more general things. But also it's very interesting because one of the things that I think you touched upon at the beginning is also a really useful like hack. And you said you've got a mastermind with ⁓ Mr. T, was it? Mr. T, the Dalai Lama and all that kind of stuff. And that's brilliant. And so...

John Bennett (56:22)

Yeah, yeah, yeah, yeah,

yeah.

Dr Dan Maggs (56:27)

I guess we're into kind of hacks rather than strategy and stuff really here. But one of the things that I really like doing, this is very specifically to YouTube videos, is I, there are a number of different sources of fantastic information out there. Podcasts are one of them, okay? And so very often people will be sharing their,

frameworks and so in the YouTube space, their frameworks for how their ideas are out of how to come up with ideas for videos or how to come up with titles for videos. And so one of the things I like doing is I will say, I'll go and get the transcript of a podcast or a YouTube video. And then basically, here's the transcript of a video. What I want to be able to do is use this

person's frameworks and ideas, write me a prompt that I can use to extract this kind of information. So I get the prompt and then I apply it to that and maybe a few more YouTube videos from that person. And then you've got an idea of their persona, their thought patterns, all distilled into this. And then maybe you can do that for two or three different people. So I've got it for my YouTube and I've got two or three different YouTube gurus with

John Bennett (57:32)

Hmm.

Dr Dan Maggs (57:51)

differing perspectives on how to do this kind of stuff. And then you'll prompt, put those personas into a YouTube, sorry, a YouTube project, a custom project, for example, and then with a set of instructions that go something like, ⁓ you are a YouTube ⁓ planning mastermind. ⁓ This is a, what do they call it? Like a,

an expert planning platform portfolio or something like that. where you basically say, so basically I want you to have a look at the personas of these people, their ideas and frameworks that are in the background in your documents and actually use those to help me give a, have a conversation. And basically you're constantly trying to outdo each other and come up with really good titles for YouTube videos that I.

might actually end up going unused. And so you get this kind of really cool kind of situation where the AI knows these people's personalities and is then able to use that effectively to out compete each other in order to give you better outcomes. And actually we've had a few really good titles come out of videos because of that.

It's really interesting and I think there is something to be said for a mastermind with the right people. I'm going to have to go away and talk about the Dalai Lama and Mr. T. as advice, but actually I think what you're touching on is actually a potentially really useful tool. And so there are little tricks like that as well. think those are... You learn those from other people, don't you? Like as you go along and if you kind of just pick up those little... tips along the way and use them as you need them. It's quite handy, isn't it, really?

John Bennett (59:48)

Hmm, absolutely. I think that's that's what's great about these sort of conversations because no matter how much you've done or what you've what you've learned yourself There's always so much more to learn in most things but particularly I think in AI, you're always finding new stuff so just just rounding off that question for Ben. So I wonder if if you could use the same approach then that you talk about there maybe find some good YouTube videos and get them transcribed and use that to help build a prompt to

maybe design the workflow for the business plan.

Dr Dan Maggs (1:00:22)

I think so, yeah. I mean, so just remember that like AI will have, you know, essentially like when we go back to what AI actually is, it's a statistical prediction engine that basically as a result of its training data is able to predict what the next token is. And so like, it doesn't know what the capital of France is,

but it knows that because of its training data, the most likely next token is Paris. And so that's why AI hallucinates, by the way, is because it doesn't always get it right. But if you think about it, AI has this huge amount of training data. And so it's gonna have some knowledge of what makes a good business plan, particularly...

you know, that common wisdom, if you like. And so that's kind of why you have sort of a tendency towards the aggregate, because actually it, one of the problems with AI is it brings you towards this kind of consensus. It's not particularly good at bringing out, you know, new ideas and genuinely new novel ideas. So you will have like genuine

John Bennett (1:01:44)

Mm.

Dr Dan Maggs (1:01:49)

knowledge from books and websites and all that kind of stuff that's been fed in as the training data as to what makes a good business plan. But also you might be interested in something that is a lesser known individual and their opinions. That would be where it would be necessarily interesting in order to actually bring in that interesting perspective.

One of the and so and then there's the idea of training data cut off. OK, and so or areas of training which it's not particularly good at. OK, and so that's when it can be really useful to pull in ⁓ much more recent kind of documents and stuff like that. And so, for example, like there's been an extremely recent update, a major update to one of the AI coding tools.

But when I talked to Claude about it, it only knows the stuff that was like a year ago, because it's training cutoff date out of date. But, so I'm having to go every time you work with this, please go and search the website. Please go and look at the documentation because it's out of date. so, it's certainly one of the problems I've come up with in the coding world is the recency issue.

And so it's basically saying, please go and look at the technical data before you give me and synthesise that before you actually give me a proper answer. But, you know, it's knowing that where the limits of it are in terms of it's got great general knowledge. You may need to give it more specific and more recent ⁓ updates as well in, you know, in your business plan example, if you've got a specific framework that you're looking for.

John Bennett (1:03:33)

Hmm.

Dr Dan Maggs (1:03:40)

for some guru that you've kind of just talked about on YouTube and stuff.

John Bennett (1:03:44)

That's great advice, Dan. Great advice. So I think maybe we've covered a lot of ground there. I think that's probably a good place for us to maybe bring it to a close, but I'd love to have you on again in the future and hear what's the latest. you... Yeah.

Dr Dan Maggs (1:03:55)

Thank you.

Well, maybe we'll make it a yearly thing and then see how

randomly different my life looks in 12 months time from the last time we spoke. Because of AI, so yeah.

John Bennett (1:04:10)

That's awesome. That's awesome.

Yeah, that's brilliant. Well, thanks again. It's been great talking to you and ⁓ I've certainly learned a lot from that and hopefully it's been useful.

Dr Dan Maggs (1:04:21)

Best of luck with it, John. Thank you so much for having me.

John Bennett (1:04:24)

Cheers Dan, thank you.