AI Chat: “I’ve got a confession to make.”
I tried going AI-first this week. Spoiler: it’s brilliant… until it isn’t.
I sat down with Nick Warren to unpack what happens when you actually try to run your daily workflow through AI.
No filters. No hype. Just two humans navigating hallucinations, rabbit holes, and the occasional win that saves a week’s work.In this chat:
Why I treat ChatGPT like an overly confident intern (mine’s called Wilson).
The moment it lied to me about Flask.
When it builds you a webpage… that looks nothing like your website.
Why voice input is finally working — and what it means for how we work next.
And yes, why Git still saves lives.
If you’re serious about working with AI (not just tweeting about it), give this a listen. (We’ll try to improve the audio next time!)
The transcript
Nick: [00:00:00] So Matt, I’ve got a confession to make. You know, last time we recorded, we were talking about AI. But when we finished, we weren’t sure whether or not to post it, because AI is such a big thing and we didn’t reach a ton of conclusions.
It’s not like the usual time we talk where—
Matt: Yeah, that’s right.
Nick: —you’re unloading your wisdom on us. But when we decided we were going to put it out, we decided to just publish the video with the transcript. And I said to you, yeah, that’s great. But I’m going to have to do quite a lot of work on the transcript because—if you’ve ever seen a transcript—from about shifting the default behaviours… And I think, you know—
Matt: Yeah.
Nick: Last time we were talking about how, you know, if you are just starting out, or maybe now you are 10 years old and obviously you’re going to grow up an AI native in the way that my son grew up a digital native. But for the rest of us who are older, or for businesses, one of the key questions we need to ask [00:01:00] ourselves when we’re doing any job—particularly if it’s essentially busy work like that—is: how should I use AI? How can I use AI to get this done quicker and move on with the things I can’t do. Does that ring any bells for you?
Matt: Yeah. Oh, absolutely. Absolutely. Do you know, it’s such an interesting point. Funny enough, yesterday or the day before, I was pulling together a list of use cases—like, how might you use AI to accelerate your daily work life?
I was Googling and going on YouTube saying, “Give me some use cases.” And you know what? There is just so much noise. On YouTube, there is so much. I’m trying to find the most influential people and what they’re saying about how to use AI.
Then there’s [00:02:00] this whole mix between how to use tools like Claude and ChatGPT versus workflow automation and stuff like that. By the end of it, my brain was just like: what the hell?
Where I’ve landed on this is taking an “AI first” approach. Everything I want to do, I’ll try with AI first. If it doesn’t improve things, I stop. If it does, then great.
Nick: Yeah.
Matt: Yesterday I had a list of tasks. One of them was to produce some copy for a welcome screen for a questionnaire survey, and then a thank-you screen. I used ChatGPT—who I’ve named Wilson. I gave him the context and he produced the copy. I added it into the system, reviewed it, moved it forward.
Then I needed to create some content for a web page about AI skills—how we’re using AI, what the skills and capabilities are. I did that too. Then I thought, okay, I’ll get it to look on the web to find out what other pages are like, and not just create the copy but the whole site.
Nick: Ah.
Matt: That wasn’t nearly as successful. [00:04:00] Clearly there’s a challenge in finding the right tool for that. Previously I’d used Vercel’s V0 tool to build web pages. This time I was just using ChatGPT.
Nick: You were generating the HTML, were you?
Matt: Yeah. I said, “This is our current website. Look at the landing pages and create a new one using the AI content.”
It wasn’t even close. It didn’t resemble our existing pages.
Nick: So—
Matt: But that’s good, because I’m understanding the limits.
Nick: Right. Because what we’re really talking about is how we’re pushing AI into the workflow. And that does mean you’re right at the edge of what it can do.
Here’s my prediction: even though AI failed yesterday, right up to the point it failed, it was very confident it could do the job.
Matt: Yeah.
Nick: That’s my experience. When you push it to the limit, it says, “No problem!” And then… you sit there for half an hour.
Matt: I—
Nick: Because it can’t actually cope yet. And sometimes it’s about context length. [00:05:00] That’s what I hit most often.
Matt: Yeah.
Nick: If you’re listening and you’re not sure what context length is: it’s how much data the LLM can process before responding. I might feed it a ton of links and say, “Let’s analyse these web pages.” But sometimes that’s just too much.
Still, I kind of like its unbridled enthusiasm up to a point.
Matt: Yeah.
Nick: And then I tell it to get its act together. Have you seen the thing OpenAI has done—where ChatGPT offers to describe you based on your conversations?
Matt: Oh yeah, I saw that on LinkedIn. Then I tried the prompt myself, and was shocked. The thing is, it feels a lot like reading a horoscope.
Nick: Right. It could apply to anyone.
Matt: Yeah. But then again, people post them on social media like it’s gospel. Did you try it?
Nick: I didn’t post it, but yeah. The prompt just popped up in the app. It basically said, “Want to know what I think of you?”
Matt: Really?
Nick: It was flattering, as you’d expect. But then I asked something more useful: “What should I work on, based on your experience of me over the past year?”
Matt: Right. Okay.
Nick: Because sure, it’s nice being told what you’re good at—but at this stage, I know that. What I want to know is where I can improve. That’s how I framed it.
Matt: That’s interesting, especially when we’re talking about confidence. AI is so confident it can do something—until it can’t. You really have to be on your toes.
For example, this morning I was using Copilot in Visual Studio Code to create a web interface. I used Python, so it picked Flask—a web framework.
It gave me instructions, but told me to run the app using python filename, which is wrong. For Flask you run it using flask run. It told me the wrong thing multiple times.
In the end I corrected it, and it updated the README file accordingly. But the point is: it was so confident.
Unless you know better—unless you know the limitations and hallucinations—you might never question it. [00:09:00] Your ability to use AI effectively is limited by your own knowledge.
Nick: Yeah.
Matt: I’ve had that loads of times, especially with code. It’s only through testing and validating that I’ve caught its mistakes. Maybe that’s less of a thing with content?
Nick: Creating text is what it’s best at—for now. But I’ve seen the same thing happen in code, spinning up a Hugo site, PHP, terminal work—stuff I don’t normally do.
Same thing with Python. There have been times I’ve asked it for something, and because I didn’t know better, I ended up wasting time going down a rabbit hole.
Matt: But that’s the thing. You might waste half an hour, but save yourself a week’s worth of effort overall. So that’s not a bad trade-off.
Nick: I think that’s true—as long as you get somewhere by the end of that half hour. There were times where I didn’t.
One example: we were messing with a Hugo template. Of course, being me, I wanted it to be slightly different from the designer’s version. I kept asking ChatGPT to make changes—“Now do this, now do that”—and eventually it got so mangled that it was easier to start again.
Matt: Yeah. That’s where best practices matter. From a development standpoint, AI is only an accelerator within good process. If it’s changing your code a lot, you need Git.
Nick: Yeah.
Matt: Because asking ChatGPT to undo changes is hard. It doesn’t track what it did well enough to roll things back. But if you’re using Git, you can revert the changes.
A lot of people using AI don’t have that developer background. They don’t know how to make their lives easier in that way.
There’s a bigger conversation here about experienced professionals versus new users, and job loss versus productivity.
There are so many peripheral tasks in every job—software development, writing, UX—that AI doesn’t cover yet. That makes it hard to replace people wholesale.
Nick: Yeah. And on that Hugo project, we did use Git the second time around—because I’d asked those questions. “How could I avoid this problem next time?”
Matt: Right.
Nick: It’s all about building a better process. Having a checklist. Writing it down. Improving it each time you hit a snag.
Eventually it becomes second nature.
One other thing I wanted to tell you—this week I’ve started talking to ChatGPT.
Matt: Really?
Nick: Yeah. You mentioned naming it Wilson. I don’t know if we talked about this before, but I think the big shift will be when we get personal AI—something that really knows us.
But for now, what I’ve been doing is hitting the dictation button on my Mac and speaking my prompts.
Matt: Oh, so you mean speech to text?
Nick: Yeah. I just tap the button, speak what I want, hit enter.
It’s part of smoothing the process. A lot of prompts aren’t programmatic. They’re casual—“I want this, not that.” And I was spending too much time typing. It’s faster to speak.
If you work somewhere you can talk out loud, try it. We’ll all be talking to our computers soon. Better to get used to it.
Matt: You know, that’s so interesting. Years ago, Microsoft did some 3D modelling thing for Xbox. The idea was gestural and voice interfaces would change UX.
In UX theory, major changes come from input shifts—like the mouse and GUI revolutionised computing.
People have talked about voice for years, but it didn’t work well—speech-to-text was rubbish. Not anymore.
Nick: Yeah.
Matt: But the challenge now is the environment. Imagine an office where everyone’s talking to their computer.
Nick: True.
Matt: You’ve got people in meetings, noise bleeding through, noise suppression not working—it’s tricky in enterprise environments.
But like you said, for those who can talk out loud, now’s the time. I hadn’t thought about this in ages, but the dream of natural language input is actually here.
Nick: I got excited about dictation 15 years ago. And again when I was writing novels. But it never worked—detection wasn’t good enough, and I’m a fast typist, so it was easier to type.
Matt: Yeah.
Nick: Now, the detection is much better. And crucially, we’re no longer asking it to create the final output. We’re asking the LLM to interpret what we meant.
Matt: And then generate the final result.
Nick: Exactly. While noise might be an issue for some, I’ve worked in offices where everyone was on the phone. So I think we’ll adapt.
You’ll be shouting to your computer from across the room while making tea: “Hey, do this.”
Matt: And it’s believable. At home, we’ve got an Apple TV and some HomePods. It recognises who’s speaking—my wife, my daughter, me. Shows an icon to indicate who gave the command.
Nick: HomePods.
Matt: That kind of detection is solid now. That’s why AI feels so hard—it’s hard to know how to use it.
Just like with remote work or early websites—people know the potential, but don’t know where to start.
Concrete examples help. Like using voice to get outcomes. Or today—I needed to write a case study for a client. We have a consistent format. I asked ChatGPT to interview me and use my answers to write it.
I didn’t need to write it at all. It just needed to extract info from me and deliver.
Nick: I agree. We ended the last episode saying people should just use it. Hopefully this one gives you another nudge in that direction. Make AI your default.
Matt: Yeah. Ask: how can I use AI to do this job? And if you’re working alone—try talking to your computer.
Nick: One trick is to imagine an eager assistant. How would you explain what you want done? That’s how to think about prompts.
For example, I needed that case study written. An assistant wouldn’t know everything—so I’d ask them to interview me. That’s what you tell AI to do.
Prompt engineering matters too. But if you can get it to ask clarifying questions, that helps.
Two tips: 1) Imagine you’re briefing an assistant. 2) Encourage clarifying questions.
Matt: Great tips.
Nick: One last thing: if you’re just using ChatGPT in a browser, try the app. On my Mac it’s right there at my fingertips.
Matt: Maybe next time we talk prompt engineering.
Nick: Yeah, I agree. It’s the first step toward levelling up—from typing random sentences to getting great results.
Matt: Cool.
Nick: We’ll speak to you next time.
Matt: Yeah, speak to you.
Nick: Thanks mate. Bye now.
