Skip to content
PodSubs
ShowsPeopleGuestsJobsNewsEvents
More
BlogPodCred
Menu
ShowsPeopleGuestsJobsNewsEventsBlogPodCredNetworksSign in
Join nowSign in
  • Shows
  • People
  • Jobs
  • News
  • More
    • Guests
    • Events
    • Blog
    • PodCred
    • Networks
    Join nowSign in

Code with Jason

Code with Jason artwork

329 - Adam Dawkins, CTO of Dragon Drop, Returns

Nov 2, 20261 hrAudio
ShareXLinkedInFacebookWhatsAppEmail

In this episode I talk with Adam Dawkins about the paradigm shift in development with AI, how expectations have changed, and the essential principles of good software design. We discuss the balance between AI-generated and manually crafted solutions, and the importance of iterative design. Links: - Dragon Drop - Adam Dawkins' Website - Nonsense Monthly

Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.

Transcript

From the show. Tap a time to play from there.

SPEAKER_00Hey, it's Jason, host of the Code with Jason podcast. You're a developer. You like to listen to podcasts. You're listening to one right now. Maybe you like to read blogs and subscribe to email newsletters and stuff like that. Keep in touch. Email newsletters are a really nice way to keep on top of what's going on in the programming world. Except they're actually not. I don't know about you, but the last thing that I want to do after a long day of staring at the screen is sit there and stare at the screen some more.

SPEAKER_00That's why I started a different kind of newsletter. It's a snail mail programming newsletter. That's right. I send an actual envelope in the mail containing a paper newsletter that you can hold in your hands. You can read it on your living room couch, at your kitchen table, in your bed, or in someone else's bed. And when they say, What are you doing in my bed? You can say, I'm reading Jason's newsletter. What does it look like?

SPEAKER_00You might wonder what you might find in this snail mail programming newsletter. You can read about all kinds of programming topics like object-oriented programming, testing, DevOps, AI. Most of it's pretty technology agnostic. You can also read about other non-programming topics like philosophy, evolutionary theory, business, marketing, economics, psychology, music, cooking, history, geology, language, culture, robotics, and farming.

SPEAKER_00The name of the newsletter is Nonsense Monthly. Here's what some of my readers are saying about it. Helmut Kobler from Los Angeles says, thanks much for sending the newsletter. I got it about a week ago and read it on my sofa. It was a totally different experience than reading it on my computer or iPad. It felt more relaxed, more meaningful, something special and out of the ordinary.

SPEAKER_00I'm sure that's what you were going for, so just wanted to let you know that you succeeded. Looking forward to more. Drew Bragg from Philadelphia says, Nonsense Monthly is the only newsletter I deliberately set aside time to read. I read a lot of great newsletters, but there's just something about receiving a piece of mail, physically opening it, and sitting down to read it on paper that is just so awesome.

SPEAKER_00Feels like a lost luxury. Chris Sonnier from Dickinson, Texas says, just finished reading my first nonsense monthly snail mail newsletter and truly enjoyed it. Something about holding a physical piece of paper that just feels good. Thank you for this. Can't wait for the next one. Dear listener, if you would like to get letters in the mail from yours truly every month, you can go sign up at nonsense monthly dot com.

SPEAKER_00That's nonsensemonthly.com. I'll say it one more time. NonsenseMonthly dot com. And now without further ado, here is today's episode.

SPEAKER_00Hey. So you were on not long ago, but for anybody who happened not to catch that episode, tell us about yourself.

SPEAKER_01I'm Adam Dawkins. I'm CTO at a company called Dragon Drop. We mainly use Rails. We build internal tools for a series of clients.

SPEAKER_00Yeah. Yeah, and we were talking pre-recording about how things have changed in certain ways now that AI exists. I've talked to a number of agency owners and people like yourself and freelance developers and stuff like that. And it's different now. The expectations have changed. Some people are using some some clients are using AI tools themselves. Tell me a little bit about that, how things have changed for you.

SPEAKER_01Yeah, it's interesting. There's sort of been this change for some clients where like quick win problems are being spun up in in like a suite of sort of single-purpose, like lovable apps or clawed apps, things like that. And it's just interesting to navigate in terms of is that even a problem? Is our job to facilitate that better? Maybe put an API in or something, or is our job to sort of um compete on speed with some of those things?

SPEAKER_00Okay. Yeah, sorry, what did you what do you mean by facilitate that better? Can you can you talk about that more?

SPEAKER_01Yeah, so they're just using like a copy of the database, right? And then they've they're getting reports out or whatever. And so I think part of our job might be to, on these like big sort of work platforms that we have, to for to sort of encourage the lovable app approach for ephemeral quick solutions or a quick different view of a thing. It's not a bad thing if someone can do that now themselves, right? Um, so maybe we could give them a better API, maybe they could have a better way to have access to those things.

SPEAKER_01Maybe it's an MCP, like who knows? Um, crucially at the moment, all of these things are read-only because they're just reading out of the database. And we're protecting the write-only stuff in the main app, but there's an interesting idea there for is this the thing that we're now building the product for? Is for AI consumption of it, or or is it more that it's just a challenge to us to use more AI in our processes such that we can build those things for them still?

SPEAKER_00Yeah, yeah. Um, and uh obviously, like you don't want to be seen as like gatekeeping or something like that. I I have no idea how much of that perception there might be out there of like clients who hire uh um freelancers and agencies and stuff like that. Um had have you seen any of that perception at all?

SPEAKER_01Yeah, it's interesting, right? Because it's that analogy of it being like the invention of the printing press, where there used to be like 20 people who could read and write things in the village, and now it's like everyone can. Like, so I guess there's the sense of, you know, and I sometimes I guess I prefer it, like when we were the only ones who could do that stuff, like, you know, it felt quite cool, I guess a bit magical. So I guess you could gatekeep that, but then like we always had a motto at Drag and Drop that it's not about the code, that that was never really the end the end goal of good software anyway.

SPEAKER_01So in that sense, I think it's it's okay to let the gates open. There's still places where good software development will still have value, I think, at least for now.

SPEAKER_00Um Yeah, and I think I'm I'm I'm really happy to hear you say that, that it's not about the code, and I'm and I'm curious uh if you could elaborate on that some more, because I think even a lot of programmers don't really grasp that fully, that when we're building software systems, we happen to be expressing the system in code, but that's not really the essence of it. We're designing a system and the conceptual framework that underlies the system and stuff like that. Um, but can you tell me more about that? What's what's behind that?

SPEAKER_00It's not about code.

SPEAKER_01It's interesting because I sort of uh I immediately feel defensive when I say it, because it doesn't mean I don't care about really good code. Um, but even really good code serves a different purpose. We talked on last time I was on about the definition of good code being code that's easy to edit or easy to change. Um and that's why that's important. I think for us it was always a focus, which was that too often the temptation for a developer is to get focused on the code and not on actually getting the solution out. Like, actually, what's the problem we need to solve the problem?

SPEAKER_01Um sometimes we're lucky enough with our clients where we have enough trust where the problem can be solved by like simplifying your time-off policy, right? Like rather than us building the algorithm that works out your complicated rules and emails alerts when people go over. What if you just gave everyone 15 days a year and it was simple? Um in a way, for me, that's still kind of probably at the edges, but that's still what we're trying to solve as software developers or software engineers. We're trying to solve problems. Um so it's not about the code for us, was that also this idea as well that we have a bit, which is that your job doesn't end when you submit your pull request.

SPEAKER_01Like that hasn't achieved we still haven't achieved anything when there's a branch. Like, if it's not live and being used and actually useful, we're we're we we haven't achieved anything. It doesn't matter how nice the code is. Um Yeah.

SPEAKER_00Yeah, I'm so glad to hear you say all those things. Um, it strikes me as a very holistic view, which is the same way that I look at it. Um like my definition of done has always been it's in production and verified to be working as expected, not when the pull request is merged or whatever. Um Yeah, and like you you mentioned that maybe you can solve a certain problem with code, or maybe it can be just some kind of tweak in the policies or something like that.

SPEAKER_00Like to me, this is an oversimplification, but like the less code that can be used in solving a problem, the better. Right. And and that like that is such a trust builder with clients, I think, because it's um well, it's like a uh fiduciary relationship where you're doing what's best for them, even if it might not be in your own best short-term interests, you don't get paid anything for them changing a policy instead of you writing code, but then they see you doing that and uh it builds trust.

SPEAKER_01Yeah, I think I I'm very lucky with our clients that we've built the trust, where we kind of get paid for the trust, I guess, as much as for the the volume of code. I think when I was freelance, it's a bit more tricky, right? To sort of stay the right side of that line, because if they could solve it in a day, but I could really do with three weeks' work, um, you get into like the mechanics dilemma of how they don't understand what's going on under the hood, so you could say whatever. Um, but I think long term that just never pays off. I feel like it's it's it's better just to try and solve the real problems and show that we can now use that other time to solve a different problem.

SPEAKER_01You know, there's always more problems, and some of them will need a lot of code. So let's work on those ones in the code, right?

SPEAKER_00Yeah, yeah. Um okay, so maybe if we can get back to the um the clients using the AI tools themselves and stuff like that. That seems tricky, and I can imagine it I can imagine it going different ways with different people. Um like I can imagine people who I've worked in the pat worked with in the past who would like make that really annoying for me. Um and I can imagine other people who like would really thrive with those kind of tools in their hands, and the whole scenario would get better, and then there's everything in between.

SPEAKER_00Um yeah, I don't know, just any any direction you want to take that.

SPEAKER_01It's interesting. S some of these what we're calling apps are sort of like nicer UX spreadsheets, right? Like like a guy would have gone and made a spreadsheet, pulled all the data out, maybe updated it once a week, and you've got a spreadsheet. Now we've got that, but it's real time and it looks nicer and it's easier to use, and you can filter it and search it. And I think with those, they probably never would have become things that like the software team would have built anyway. Like they're low enough hanging fruit that we've just improved the productivity at the low end because we just have a better version of a thing.

SPEAKER_01Um there's a tension on speed because the idea it is a bit uncomfortable when it feels like your client's faster than you at delivering the the thing that they couldn't do that, right? Um and they'll probably listen to this as well, so I'll uh but but it's okay because I think it's like again, it wasn't about the code in the first place, right? What what's interesting is um I think there are some risks to it. Like I I don't think um the speed difference isn't because we're typing slower than lovable can type.

SPEAKER_00Right.

SPEAKER_01Right? The speed difference is the test suite, the A3 sheet of paper with the flow diagram, making sure it connects with everything else, putting the authorization policies on top so the users who can't see it can see it, all of that sort of stuff. That's the that's the extra time. You're actually getting more. Like it's it's not um a fair comparison of of time.

SPEAKER_00Um Yes, I think you're right. I want to maybe challenge that a little bit and add to it. Um you know, some of the speed that you get with with uh agentic coding is an illusion. You know, it's like you're you're you're getting ninety-five percent of it immediately, uh but the other crucial five percent that you absolutely need uh is i the it's like kind of painting you into a corner so that that last five percent is gonna be almost impossible to achieve.

SPEAKER_00And so the the apparent slower speed is um doing all the work in such a way that at all times you have a tidy foundation on which to build the next thing rather than painting yourself into a corner. I think everything you said is true, and then I wanted to add that also.

SPEAKER_01Yeah, yeah, I think it's a different conversation almost. Developers using agentic code is a is a completely different thing to sort of vibe coding solutions. They where you're just asking for an app that does X and it's like a one-prompt, one-shot thing. I think that's very different to how we should use agentic code to like speed up our processes or whatever. They feel like completely different things. Good point. Like the way I approach it with Claude is I actually get Claude to commit a lot because then I've just got all of these tiny stage gates of decisions, and I can just we can just go back three steps rather than coming to the end, and Claude says that's done, and it's like plus 900, minus 800, and you go, well, I hope it works, because I've got no way my next reset point is to start the game again, right?

SPEAKER_00Like a reset point.

SPEAKER_01Um so I think that's a slightly different thing. I think this is more Yeah, it reminds me of that that uh printing press thing a bit, because it's kind of like if I wanted to read something, I had to ask the people who could read. And now I can just read it myself and it's it's done. I don't need you for that. Um so that's like a really interesting analogy, and I I and I don't think we should see it as a bad thing, because if you're a software developer, you probably believe that software's good, and if there's more software, that's probably a good thing.

SPEAKER_01It's just interesting that there are parts of the craft that that aren't known to those things necessarily.

SPEAKER_00Right? Yeah, I'm okay. I have I I'm not totally comfortable with the printing press uh reading-writing analogy um because literacy is almost like you have it or you don't. Um you know, there's degrees, but in in a sense it's binary also. It's almost like if everybody could suddenly just have all the types of power tools and building supplies and stuff like that, lumber, just at at a snap of a finger.

SPEAKER_00And and then you can say, Oh, okay, I'm gonna build a wall here and a whatever here, electrical system, blah blah blah. Um because there's okay, with reading and writing, there's a lot less room to shoot yourself in the foot. Sure. Um whereas with programming there there still is that.

SPEAKER_01Right. It's like I could uh I could clog myself a house really quickly, and the builders would take longer, but their one doesn't fall down when I shut a door, right? Right. Um I'm not gonna electrocute myself in the house that the builders build. Whereas I might in the sort of AI generated version. Because I don't want to.

SPEAKER_00And it's funny because I'm second guessing this analogy as we're talking about it, because like uh AI probably could design a pretty good house, I imagine, a lot better than a lot of builders, but maybe we can leave that.

SPEAKER_01Well, I think no, but I think the weakness is maybe, but I think the word could is the whole issue, right? Is that it's non-deterministic and it's not guaranteed. That's the whole problem. It's like sometimes you roll the dice and you get exactly what you needed, but other times you don't. And if you don't understand what the dice roll means, how do you know which one you got? Right? Like, that's the bit I find interesting is how do you know that you got bad code back if you don't know what the code means?

SPEAKER_00Well, there's okay, so when you used the example of the spreadsheet, that was an interesting example to me, because um, you know, spreadsheets are two-dimensional mostly, you know. Um but then a web application will usually be um backed by a database and you can kind of you can you can take steps from a spreadsheet uh data storage approach, data manipulation approach, to database based approach.

SPEAKER_00Um and there are natural laws behind that. Like if you have okay, I have a spreadsheet of contacts. Um like I I I really do. Um everybody I know and their contact information and stuff like that. Um one of the rows in the database is sorry, one of the one of the columns in the spreadsheet is location. Uh and I might put like California, Los Angeles, California, San Francisco, so that when I sort it, it's uh kind of sorted by region.

SPEAKER_00Um if I were to put this in a database, um l let's let's say I'm a non-technical person, it might be clear to me that I should make one row per uh customer. Sorry. One one row per contact. Um but because I don't know anything about relational theory yet, um I will happily stick that location value in each contact uh record, unknowingly creating data duplication, which might bite me down the road, or at very least cause me great inconvenience when I want to do certain kind of queries and stuff like that.

SPEAKER_00So I'm just kind of thinking through this out loud right now. Um it's it's maybe different from like building a house and stuff like that, in that they're right out of the gate, you run into these like natural facts of reality. You know what I mean?

SPEAKER_01Couple of points on that. It reminds me of Airtable a bit, the problem you described. We did some consultancy for clients on Airtable, and Airtable's like relational databases that are spreadsheets. So you kind of get typed columns and you get links, but it's otherwise still spreadsheety-y. And we found that the biggest problem wasn't the tool, it's really easy to use, but you have to know when something's a different table, right? And that's like actually very hard to understand because sometimes it's not very intuitive. Like we understand relational databases and they seem intuitive to us, but it's not always clear what isn't just another column and when something is a different table, and things like that.

SPEAKER_01But I think my analogy for spreadsheets was more this idea of they've got the data in the app and they just want another view of that data, like a particular way of looking at it, right? So I'll just do a quick spreadsheet that groups everything by this. It's kind of the um, I was more relating to spreadsheets in the sense of they're sort of um not necessarily long-living. Your contacts one sounds like it is, but just like a quick bit of research to show some data in a certain way. Oh, okay. I can give that to some people and they can see it, except now it's a web app instead, right?

SPEAKER_01Um, which is another one of my issues a bit with AI, is that it it's very good at looking like a really good thing. Right. It's another issue, like if someone sends you a report, reports used to look good because if someone has formatted all of the headings, they've probably spent a lot of time thinking about it, right? And now it's like the level of polish is no indication of the level of confidence you can have in the thing. Um, like if you sent me a web app you'd built, if it if the UI's really good, I know we're quite far down the road, probably, because you probably did that bit last, it kind of indicates.

SPEAKER_01A certain thought process that went on. Whereas now it's like the uncanny valley of the AI just gives you the thing that looks like the real thing straight away.

SPEAKER_00Yeah. Yeah, and I've heard of people, and I've adopted this practice myself of like don't show clients like really high fidelity mock-ups, because then they will very understandably get the idea that it's like almost done. It's like all you have to do is wired up to the database and make it work. Like the the hard part's already done. Now you just have to make it work, you know? When really that's obviously backwards.

SPEAKER_01Yeah, when I switched to doing more back-end work, I started as a front-end developer. Uh and when I switched to more backend, that was always the frustration was I've done the really hard bit, but like someone's redesigned, restyled a page, and everyone's talking about like the massive amount of work that was done on that page, and it was like we just changed some classes, you know. Um, I think the other reason there is it the other reason I agree with you to not show high fidelity is that actually it goes back to what we were saying about it's not about the code of what is the problem we're trying to solve.

SPEAKER_01And that once we've got a whole UI in here, we're gonna start thinking about colours and borders and button locations when actually I want you focused on the problem. So sometimes I'll I'll even open Vim occasionally on a client call and just type with arrows, just to this is the talk, help let's understand this. This is all we need to solve right now. Um, and that's the foundational bit, right? Like I think it very rarely have I solved a problem well where there weren't some arrows written down somewhere between some things, and then we built on top of those, assuming that those have like that hypothesis has been proven so we can move forward to the next layer of fidelity.

SPEAKER_00Yeah, yeah, it's astonishingly rare to go through a planning process and and have the people involved uh state the essence of what's trying to be done, um not at too low of a level of detail where it misses the point, and not at too high of a level where it's too vague, but getting it at just the right level and hitting the problem on the head. Because that that stage is when you have so much leverage and influence over how much the rest of the project costs relative to what it should.

SPEAKER_00Like if you get that planning part wrong, the rest of it is gonna be um so problematic, and if you get it right, then um you save so much work. So I applaud your approach because I I've I've been on a lot of projects where a lot of time is spent like designing the UI and stuff like that. Um, but it's like hang on, like what what actually are we trying to do? And it's so easy to get into discussions about side topics and stuff like that, that really uh really doesn't even matter.

SPEAKER_00So yeah, I agree with that approach.

SPEAKER_01So when you're working with internal tools, the UI is is fairly solved, I think, because there's only a handful of patterns for capturing data, displaying data in the right way. UI design's probably a bit different when you're doing something public-facing for a brand, because there are other concerns there, and that art form needs to be applied to those other things. But I still just think more generally, I don't know, I think I approach a problem like a like a hypothesis. Like we never know less about this problem than we do right now, right?

SPEAKER_01And so sometimes it's like, let's try and let's stay on pen and paper until we can't learn anything else from pen and paper, then we move up to the next level of fidelity until we can't learn anything else in the console or whatever, and then eventually we'll build the whole thing.

SPEAKER_00Um, I think this is I okay, so I'm trying to pay you a compliment, not not be critical, but I think this is um a case of unconscious competence where like to you it seems like there's an obvious path forward for these internal tools, but there is so much room to get it spectacularly wrong. Um and man, your approach is so like refreshingly rational.

SPEAKER_00Um let's let's start with pen and paper and stick with pen and paper until we can't do that anymore. That that is like I think that overlaps a lot with the way I think, but it um is maybe not quite the same. I'm curious to hear more about that. Like why why that approach?

SPEAKER_01Yeah, well, pen and paper's probably a bit loose. Sometimes it's Venom typing, but it's the same idea. It's just too um I don't know, I want to poke at the problem at the simplest point. And it's it's it's that the idea is that pen and paper is the easiest to throw away, right? And and maybe this is where AI is changing things a bit, because you can build three full solutions maybe and pick one, I don't know. But it's kind of let's iterate at the cheapest level of learning and being wrong. And so with pen and paper, I you know, not bragging, I can afford more paper, right?

SPEAKER_01So I just throw it away, use another sheet, we'll start again, big cross through it. Uh, I use a massive Sharpie for it and just like do that. And then it's like, okay, little prototype. Again, we can we can throw it away at any point, but we'll we're based on the learning that we've confirmed at a certain point.

unknownYeah.

SPEAKER_01Yeah. And and the you were right earlier. I think you said that the biggest it doesn't matter how far we build if the thing we're solving isn't the real problem. It will be wrong. And we will waste money and time, and everyone will be frustrated, and and we haven't achieved anything that was all solvable on the paper. None of that needed to that all could have been discovered at the discussion phase.

SPEAKER_00Right. That's an excellent point. And I guess I'll also add that um when you're working in pen and paper, there's no there there's such a smaller chance of fooling yourself into thinking you're doing something productive when you're not. Like there's no UI that you can see and trick yourself into thinking that uh that you're creating value when you're not and stuff like that. So I like that aspect about it also. Um because I do the same thing for similar reasons.

SPEAKER_00Um like I start on on pen and paper because like if you're on pen and paper and you're wrong, like there's there's nowhere to hide. You know, you can't hide your mistake behind anything. And like even more essentially, I think, your point about the um fast iteration and feedback. It's cheap to do it over. Um this touches on another like fundamental aspect of reality.

SPEAKER_00There's this book that I talk about all the time, The Beginning of Infinity by David Deutsch. Um and he talks in the book a lot about Karl Popper's uh conjectures and refutations and how science advances by scientists coming up with hypotheses and then subjecting those hypotheses to scrutiny, both their own scrutiny and the scrutiny of other people. Experiments are performed to subject those hypotheses to the scrutiny of reality.

SPEAKER_00Um and we're doing the same exact thing as programmers. We um you know, every design is a hypothesis. I I think that's what you were saying. Or less. Um and then we subject these designs to scrutiny, both our own scrutiny and the scrutiny of others. And usually, just like with science, our early hypotheses will be very uh I was gonna say wrong, but the wrong isn't exactly the right word, I I think. But they'll they'll be off the mark in some way. And so we'll we'll iterate, and as we gain confident confidence that our hypotheses are on the right track, we can move up to more and more expensive higher fidelity uh forms of the hypothesis, being increasingly sure that we're not wasting our resources.

SPEAKER_01Yeah. Yeah, I think that's broadly it. There's an interesting thing this touches on for me when we talk about this, is that with AI coding, one of the problems I'm coming up against is that I've realized that I trust my process more than I necessarily trust the code I end up with. Does that make sense? Like I trust how I arrived at it. Like if I'm if I'm doing a big refactor, I trust that every that the tests passed at every step, and I knew that I could delete that safely, then delete that safely, then change that safely.

SPEAKER_01But if you just showed me my own code at the end, I don't know if it's right, because it's just deleted a bunch of code. What if it hasn't deleted the other bit that it needed to delete or hasn't, whatever? And so that's a real challenge for me if I ask Claude refactor this and I've trained it to use all of my processes or whatever, is it just gives me the answer. But I found that one of my tensions is that a lot of the time I'm not relying on my ability to read code as much as I'm relying on trusting my TDD method, trusting my approach. And that's really hard to do once you switch to the agent because I don't know what approach is followed.

SPEAKER_00Okay. I'm curious how similar your um agentic coding approach is to mine. So I've tried to take these like principles that apply no matter what and build them into the agent. Um so I have this like TDD skill where I say, and I've talked about this several times on the podcast before, where where I say that I I give it some requirements, it repeats back to me in its own words what I said, then it takes one of the so some one-line specification from from what we're trying to build, translates that into a test, runs it, makes sure it fails in the expected way, um, then writes the code to make it pass, and it's checking with me every step along the way to be like, does this look good to you?

SPEAKER_00Um and then, you know, once once we get each test passing, we commit, committing very frequently, just like you said you do. Um and I found that that it's it's hard to go off the rails that way because um again the principles underlying this way of working are just sound. And so if you follow these principles, it's hard to go wrong. But I'm curious if you do some if you if you do it that way or some other way.

SPEAKER_01Yeah, that's broadly what I follow. I haven't got it quite as strict, um, although it does keep it's finally learned. I've got a cloy, it's like writing down memory of me every time I say no. So it's starting to write the tests before. It went through a phase of saying that it was doing a strict TDD and it would write all the tests at the end, um, which was very frustrating. Um but then it like but it would write like the last commit would be like TDD steps for, and then it would just be like the all the tests at the end. I I just wonder if there's something different where it's like I guess I trust myself behind the wheel, right?

SPEAKER_01That's the challenge. And it's it uh and I think it's um it's easier to use it for something I don't know how to do. That's easy, because I don't know how to do it, so I can follow that through. But it's like if I if I know I can get there in an hour, Cord might get me there in five minutes. It might also still be wrong in an hour, and there's that kind of confidence of I know if I do it, it will be done right, and it will only take an hour. It's sometimes hard to I sort of feel this almost obligation that I should be using it when sometimes I really don't want to.

SPEAKER_01I just want to type this, I know what I'm doing, you know. Um but some of that might even enjoyment as well. Maybe I just enjoy coding, so I don't really want to use the other thing.

SPEAKER_00Well, I I can speak for myself. I feel like I'm still in a calibration period where I don't know I don't know where to set all the dials, you know. Like uh I I've been working on this thing over the last several business days. Um at at my job we're implementing um badged merging. So I'll I'll try to keep this background brief, but we have like 2,000 engineers.

SPEAKER_00Um every day we have like a hundred pull requests. Um and right now the the pull requests are just being merged whenever they pass the tests. The problem is we can get what we call soft conflicts where there are two changes which are fine independently, they they pass the tests, but once those changes are together uh in the same branch, they interfere with each other and cause a bug. We wanna we wanna prevent this because uh currently we can get these bugs on the main branch and then it causes all sorts of headaches, and you know, the further downstream a bug gets, the more expensive it is to correct.

SPEAKER_00Um so one way to deal with this is instead of testing things uh individually in parallel, we test them in batches serially. So maybe if we have a hundred pull requests a day, we do five batches of twenty, for example. Um but what I don't know is like how exactly is all this gonna play out? Um like how frequently can we expect a bad batch? And then when we have a bad batch, how what's it gonna cost in terms of time to winnow out those bad uh PRs and how long's all that gonna take and stuff like that?

SPEAKER_00So it and I'm I've been tasked with writing a document to explain all this to everybody and uh tell them what they can expect and stuff like that. I'm like, well, I can't tell them what to expect before uh I have some understanding of what's gonna happen, so I decided to model this out. Okay, all that background is to say I've I've been working on this small one-page JavaScript application where you can plug in some certain variables, like how many PRs a day, how big are the batches, etc.

SPEAKER_00How long does one batch take? And it'll give you a table with with rows saying here's what all the batches shook out to be, and here's how long this thing took in total. It took 17 hours to get through everything or whatever. Um the first like 90% of it was really easy, and then the last like 10% was just like, oh wow, it really painted me into a corner, and now I need to go back and do a bunch of refactoring so that I can even conceive of how to make this last change that I need to make.

SPEAKER_00Um and it's like, did I w was I too loose, you know, with with my um did I give it too much rope? You know, did I give it too long of a leash? Uh there's an argument to be made that yes, I I I didn't really look at the code that much because I'm like, this is a one-off throwaway application, we're gonna use it this week and then never again. Um, so it doesn't matter if the code is bad, but then we got a certain depth in, and it did matter that the code was bad.

SPEAKER_00So I'm still in this like calibration period. Okay, sorry, it seems to have happened again. Uh I said I'm in this calibration period, and then I didn't hear anything from your end after that.

SPEAKER_01Uh yeah, I think I got a bit more than that. Um I got up to like w it was gonna be throwaway, so it didn't matter, but then we got stuck.

SPEAKER_00Um, yeah, and then at the end I said again, so I'm still in this calibration period. Um But yeah, basically just what I was saying was like um the the code was too sloppy, painted myself into a corner, and so I'm like, okay, maybe that wasn't the way that I should have interfaced with the agent. So I'm I'm still in this period where I'm experimenting and trying to figure out what's the best way to behave and interact with this tool.

SPEAKER_01Yeah, I wonder if I could play an example back to you and we'll see how it compares. Um so I had to solve this problem. We talked about it before the before the recording, of um the idea of how do we how do we bulk assign sales reps to appointments for minimal drive time? And so I I uh Claude helps me find out that the thing I want is the Hungarian algorithm. This is the linear sum assignment problem, and that's what that is, right? And I feel like that's a great use of Claude where I sort of use it where I would have used Google, because that's like tons of reading, and oh, is this the right one?

SPEAKER_01Is that the right one? You know. So then I asked Claude for a Ruby version of the Hungarian algorithm, and it gives me one. And then I say, Can you prove that it works? And it created um uh a brute force version, which was like two lines of Ruby, whereas the main one was like 150 lines. Because the two lines of Ruby essentially just um this algorithm takes a cost matrix, right? So you've got like people down one side, tasks across the other, and there's a cross cost matrix of how much it would cost to do each thing.

SPEAKER_01In our case it was in minutes, but it could be in anything, right? And it just picks the optimal cell for each row, and that's how it how it does it. But the the brute force version just used the Ruby permutations method, which just creates all of the combinations. So it was great on a on a two by two, but once you get to 20 by 20, there's like quintillion it just doesn't work. So then I said, How can you prove this at the big numbers? I want 20 by 20 and I want to see that it works. And so it downloaded and installed, I was spiking, so I was letting Claude run, and it downloaded and installed the SciPy version of the algorithm from the Python library and put the same example in that and compared the results.

SPEAKER_01Um, which was okay, but it only compared that the sum was the same, not that the each selection was the same, right? So so what we decided to do was keep Claude's version of the algorithm. It would have been fun for me to write it myself. That would have been a great two days for me. There's steps on like Chicago University website or something, and it's on Wikipedia of how you produce this algorithm. But I've got this thing. So we worked on finding boundary tests for that thing that gave us some confidence that it worked.

SPEAKER_01And then we actually went into the SciPy C library and copied some of its tests into our tests and checked that it still worked. And then I put a performance test on because with the example of a 3x3, you could still collapse all the code down for this really simple thing, but then it won't skip, it will blow up if you use it with big numbers. So that bit was clauded, but that bit doesn't produce the matrix. So we've now got to produce the matrix from our database of appointments and reps and product types and appointment types and all of this different stuff and giving things big numbers when they should be, they shouldn't be able to run it or whatever.

SPEAKER_01So I built that really manually because I wanted that was the crux of the problem. But while I was building that manually, I realized I needed a really simple predicate method on on my user model. So I just asked Call to go and write that. And it wrote the test as well, and it's like two lines, one line each side, right? I can check it really easily. And then I and then and then for the last step, we're just throwing this together on a page, and I just wrote some of the ugliest code I've ever written, just like let's test this thing out. It's just all in like instance variables, we just pull them all out on the page.

SPEAKER_01The controller action's like 30 lines, it doesn't matter. So I feel like I got a good balance there where I used Claude for the hard bit, but then really slowed down to verify that that was the right thing, and then use Claude for the really easy bit, the other side. And then Claude's not much faster than just writing really terrible code yourself when you want to do that as well. So that was kind of the tension we tried to strike there.

SPEAKER_00Yeah, interesting. I've I've found the same thing where like Claude will clue me into something that I didn't know about. Um, and it's the best when it's something I didn't know I didn't know. And so it's like something I wouldn't even have looked for, but it's like, hey, you realize this thing you're doing, it like kind of pattern matches for this like very solved problem. Uh do you want to just use this thing? Um and that that has helped me a lot. Um Yeah, yeah, it's really interesting to um figure out where it it shines and where it doesn't.

SPEAKER_00And I think that um there's a great there's a great variance right now in how well and appropriately agentic coding is being used. Used. Um I I see a lot of developers just like uh taking a a ticket and pointing the AI at it and say like have at it. Um that doesn't seem like the way to go. Um and yeah, I I don't know.

SPEAKER_00And and are these things that people will figure out over time. I'm not optimistic that that will be the case, because like there's this kind of um group of people who have like figured out uh this this this process of iterative design, starting with very lo-fi designs, um, and then and and then feeding that into the TDD process, which I see as just a continuation of the same idea.

SPEAKER_00And I feel like these things are like again, naturally true facts. Like people independently discover these same ideas because they just um they have merit uh no matter what. And before AI, like a very only a very small fraction of developers are working this way. And so my like kind of cynical expectation is that uh there will still be like 90% of developers working in this uh less effective way.

SPEAKER_00I I don't know, I don't have a point or anything. I'm just thinking of it.

SPEAKER_01So it's interesting though, right? I sometimes want TDD is the classic example of I trust my process more than my output, right? I trust my PR at the end of it because I know that I followed this approach to get there. And that's why when I review other people's code, it's it's more exhausting than looking at something I get to at the end because I know how I got there, right? And I sometimes wonder, even when I get clawed to TDD, if it's just cosplaying at it. It's already decided everything it's gonna do. Whereas the value of TDD as a design tool is that at each step you then consider, you know, it's it's red-green refactored.

SPEAKER_01This is where I consider if this design's still going in the right direction. Probably one of the underrated parts of TDD, maybe, is that it's forcing me to go in really small steps. And at each step, I'm deciding consciously or subconsciously whether I'm gonna change my mind, whether this design is still holding, whether the five steps I know that are coming up are still gonna make sense. And I wonder if when I say to Claude, do the whole thing TDD it, if it just works out what the whole thing will be and then puts it in the order I've asked for so that I think it did TDD. But that's not the same as like test-driven design, right?

SPEAKER_01Like if you want to call it that as part of test-driven development, that's just writing a test, then writing a writing some code. It's kind of a commit order more than the actual process of letting the tests let letting the design emerge from a series of tests, right?

SPEAKER_00I think that's okay though, because to me, uh maybe the number one benefit of test-driven development is that it results in full test coverage, and when you have full test coverage, you have the freedom to refactor liberally. And to me, that's the main benefit. And so I don't really I don't really care how the I mean I do care how the test came into existence. That that matters. Um But even if if Claude like predetermined all of it somehow and still wrote the same thing, if I still have the benefit of being able to refactor, that's the part that's really important to me.

SPEAKER_01Yeah, that makes sense. Yeah, it's a very interesting space at the moment. I was even thinking with what I described, I did with that algorithm, that actually the the bit that was the most developer-y that I did was asking for an algorithm. Right? Like, what's is there a known algorithm that solves this type of problem? And I think if you just do a one-prompt, like a one-shot prompt to Claude or Lovable or any of these things, you you you might not even get to an algorithm that's gonna deterministically produce the same results again and again.

SPEAKER_01And then it's like even that sentence, realizing that in this context we need something deterministic, is still like the skill of the software developer, right?

SPEAKER_00Right, yes. Yeah. And I wonder if there's gonna be like this new um kind of uh uh professional skill where you're like a partway developer. Um because I don't think it's all or nothing. You know, I I think it's just and I don't think there's even like an inflection point or something like that. It's like if you want to become an electrician, you like go uh you you apprentice or whatever and then get a license.

SPEAKER_00I don't know how it works, but there's a license at some point, you know? There's a binary, you are not an electrician yesterday, and then today you are an electrician. But with programming, it's it's a continuous spectrum. Um and I can see people who aren't developers learning some of these like basic natural truths so that they they can build these programs without shooting themselves in the foot.

SPEAKER_01Yeah, my analogy to that would be if I was to work in programming language I don't normally work in, right? You give me like the knowledge of the principles plus Claude and copilot in the editor, like Go syntax isn't gonna trick me up very often, right? Because the AI is gonna solve the syntax issues, and I'm gonna solve the software design issues, right? So so you could see it being more like that, that if you learn how to design systems, you might be in a good in a good place. Um but I don't know, it's hard being a sort of quite an I mean quite an old developer now.

SPEAKER_01It's difficult to know how would I how do you get to this place without just like thousands of different examples going wrong on you over a period of years. That's kind of how I built up the knowledge. I didn't really have like a syllabus, right? It's like I'm sick of this problem happening, there must be a solution to it, or um that intuition that develops over time. I don't know how you develop intuition if you're only agentic coding in your whole career. I don't know. I'm not saying it's not possible because I don't want to be that guy who says the next thing is impossible, but how how does that intuition develop where you look at some code and you just think, hmm, there's something wrong here.

SPEAKER_01I don't know what it is, but it's like just your brain has seen these patterns so many times, it's recognizing it subconsciously. I don't know how you get to that when it's so heavily automated.

SPEAKER_00Yeah, that's a good question. I I've found that that exact phenomenon that you're describing has happened to me. Um I've also found that it has like heightened other senses of mine. Um like things that I never really thought about that much before because it's just the way that I do things. Uh Claude will do certain things very differently, and in my opinion, worse.

SPEAKER_00And I'd be like, oh whoa, whoa, whoa. Like, okay, we're not doing that. Uh, but it's it's something that I've never really thought about before. Here's here's not a very good example, but like heavily commenting code. Like, it just wouldn't even occur to me to think about that because I never do it, but then I ask Claude to write something and it puts all these comments, and I'm like, okay, that's going in the skills file. Like, don't don't comment the code redundantly saying what's what's happening, don't do that.

SPEAKER_00So I I've experienced both.

SPEAKER_01This is a fascinating area for me. I've been speaking to some friends even across like different industries of when you manage a team, of the default is do it the way I do it. And because I know the way I do it works, but the challenge I think is, and it's probably a whole different conversation. The challenge is to work out the principles within what you do that are what work, right? Uh, because not everything you do is what works. Some of what I do is just weird. Like I use a Dvorak keyboard and I use Vim.

SPEAKER_01None of those are essential to the workflow, although I would fight you for them. They're not essential, right? Like those are just quirks of how I do things. And so what I find interesting is how do you how do you identify what are, and maybe Claude forces you to do this a little bit as like a mirror, how do I identify the principles in what I do that makes what I do consistently come out right for the last 10-15 years? Versus, well, I've always done all of these things this way, so so I I you you have to take all of it or none of it, you know, like when you're training someone else or when you're showing someone else.

SPEAKER_01Um, because probably the thought under the thought with the not writing loads of comments is we're gonna write code that's clear enough that it doesn't need that much commenting, right? And so the maybe the principle is more about clear code uh or or something. It's just really, really interesting. And a few times lately I found myself questioning and I've apologized on code review to some of my team, where I've said, I'm not saying it's wrong, I'm just saying I've never needed to do that. And that's quite lazy for me because I'm leaning on just experience, but it's also got a fair chance of being right because I have the experience, right?

SPEAKER_00Um, yeah. Man, I'm I'm gonna send you the manuscript of this book I'm working on, and and you can glance at it if you want to. Um, but it talks about this exact thing. In fact, the whole point of the book is this exact thing where it's like, how do we distinguish between what is just a personal preference and what has like a a natural necessary truth to it?

SPEAKER_00Yeah, and like for example, um, I'm a testing chauvinist, you know. I I would never consider not testing. I think testing is objectively better and smarter than not testing, but it is that just a matter of opinion? Um and if it is, why? If it's not, why? That kind of thing. And it might seem like just um just um uh pointless, unproductive philosophizing, but I don't think it is.

SPEAKER_00Uh I d I I think it's really productive to get on a firm footing regarding um whether these things are objectively true or just opinion.

SPEAKER_01Yeah, I think so, because even if then well this is gonna sound weird, but even if you're disagreed with, we understand what the trade-off is, right? In that TDD definitely gives you something. There's there's some things it definitely gives you that someone who hates doing it can't argue with. Now, whether that means you can convince them to do it might be a different thing, but we can at least recognize the trade-offs. I heard Kent Beck talking about that when he said, well, we found that we only had bugs when we didn't TDD and we didn't pair program.

SPEAKER_01And then he said, You can do what you want, but but you can have the bugs then. We didn't get bugs when we did it this way, you know. Um, which is yeah, slightly different. I I I just I I I I think it's a really useful exercise. I just wonder if people will ever fully agree with what conclusions you come to, even if you've like tried to prove them, you know.

SPEAKER_00Yeah, yeah, and that's okay. Somebody asked me, like, what's my intended audience for this book? And I'm just like, the intended audience is me. Like, I'm just writing it for myself, and I'm I'm betting that if I'm interested in this, maybe a few other people will be also. But I don't care if anybody buys the ideas. Realistically, a lot of people won't. Sure.

SPEAKER_01Yeah, it's interesting. I think it it's a weird craft, right, software development, because all the other examples we talked about, like engineer or electrician or whatever, you I think I've heard you talk about this before. Like the principles are established in a non-negotiable, right? Well, like, like there's a way you can build a bridge. There's like, I don't know how many ways, there's maybe five ways you can build a bridge, but gravity's gonna win against all the other crazy ideas you come up with, right? Whereas encoding, especially in, like as I said, I use Rails a lot where you're basically allowed to access anything from anywhere ever at any point, right?

SPEAKER_01Like you can write it in the most horrific way possible, but there's not like there are principles like solid and whatever, but they're not the same as like the certified standards of engineering or or of civil engineering or those types of things. You know, we don't really have that in our communities, I don't think.

SPEAKER_00Right. Yeah, the feedback that we get is much more mixed in with um other factors. Like with a bridge, the feedback you get is very binary. Uh is it still standing 50 years later or is it not? Um Whereas with the software program, what constitutes a success? You know, like uh it's it's often said that like if you have a really messy legacy application, it's 20 years old and it's a piece of garbage, that's a great success because that's supporting a business that has succeeded for 20 years.

SPEAKER_00It's like okay, yes, but also like what if the code wasn't a mess though? Like, how much more successful would that business be? So I I I I'm I'm skeptical of of that argument. But I I think the explanation behind that is just there's so much stuff mixed in. It's like more akin to like um psychology and stuff like that. Like, how do you successfully raise a child? Right. Well, what's the feedback on that, you know?

SPEAKER_01I guess part of it as well might be that um I mean there could be good, there can be terrible consequences to bad software, but it's never like one of the principles, right? There's always a whole multitude of things that went wrong, or at least more than one. I guess there's something where if we if you build a bridge and then a car goes on it and it breaks, that's really bad. Whereas like if the first user hits this page, it goes down, we'll just fix it. Like there's kind of that um we are we we do produce text for a living still, like, which is why AI is so involved in coding, because it's a it produces text, right?

SPEAKER_01So I guess there might that might be part of why is that the iteration cycle is smaller than say re like an yeah, I'll call it real engineering, you know. Um it it it there's maybe some of that, but it's just interesting where it's like it feels like every principle you want to apply can be a debate in the wrong crowd because nothing's sort of universally agreed. Um maybe that's okay. Maybe that's what makes it interesting. I don't know.

SPEAKER_00Well, I feel like there's a lot of lost productivity. There's having arguments, and it it's it's not necessarily what the arguments are about, it is the quality of the argument um where it's just my opinion versus your opinion. And that's that's not that productive because usually nobody comes away particularly moved. Yeah. Um but if we can shift that and say this isn't just my opinion, you know, one plus one equals two, and let's build up from that foundation, then I think that can be a lot more productive.

SPEAKER_01Yeah. Yeah. I just don't know if we'll get there.

SPEAKER_00Hopefully someday. Um, okay, we should probably start uh wrapping up. Um Adam, what uh what links and all that would you like to share for people?

SPEAKER_01Um same really dragandrop.uk is the business, and then um adamdawkins.com, I believe, is my website. It's very infrequently updated. You can email me at adam at dragandrop.uk.

SPEAKER_00All right. Well, Adam, thanks so much for coming on the show. Thank you.

Open the transcript file

Hosts and Guests

Nobody is credited on this episode yet.

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. 338 - Code Review, Testing, and Staying Sharp in the AI Era with Joel Hawksley

    Nov 18, 20261 hr 5 minAudio

    In this episode I talk with Joel Hawksley, Staff Software Engineer at GitHub and creator of ViewComponent, about whether AI is making us dumber, why blocking code reviews struggle under a flood of AI-generated code, and...

    Credit the people on this episode

  2. 337 - User Testing, Paper Prototypes, and Escaping Hourly Billing with John Athayde

    Nov 17, 20261 hr 18 minAudio

    In this episode I talk with John Athayde, designer-developer and co-author of The Rails View, about what architecture school taught him about software design. John tells the story of a user test at LivingSocial that...

    Credit the people on this episode

  3. 336 - Riccardo Carlesso, Co-Organizer of Rubycon Italy

    Nov 13, 202646 minAudio

    In this episode I talk with Riccardo Carlesso, Google developer advocate and co-founder of Rubycon, a Ruby conference held in Rimini, Italy in May 2026. Riccardo explains how a joke about PyCon became the conference's...

    Credit the people on this episode

  4. 335 - Blastoff Rails: Starting a Ruby Conference from Scratch with Travis Dockter

    Nov 11, 202654 minAudio

    In this episode I talk with Travis Dockter, organizer of Blastoff Rails in Albuquerque. When Sin City Ruby and RailsConf both ended, Travis decided to start his own conference. We discuss how cheaply you can run a...

    Credit the people on this episode

  5. 334 - Merge Queues and Building Mergify with Julien Danjou

    Nov 10, 202655 minAudio

    In this episode I talk with Julien Danjou, CEO of Mergify, about merge queues: why PRs that pass CI separately can still break main together, and how testing changes in a queue keeps main green. We discuss how Mergify...

    Credit the people on this episode