Codurance AI Hackathon

This is quite a long post (15 mins) and if you’re only interested in my final musings, and not the day itself, just skip to the last two sections.

Last Saturday I got to attend an event at Codurance’s offices in London around the use of AI based tooling to aid software delivery. I did not have much experience with these kinds of tools, in part because my current client uses their own in-house programming language [1]. Hence, my experiences to date have been limited to some messing around in VS Code with Copilot on some fairly simple C# katas.

Event Format

There were about twenty of us spread across a range of ages but, generally speaking, all fairly experienced programmers. We were split into two teams (A and B) and got to tackle two problems – one in the morning, and another in the afternoon. We got to spend 2 hours on each problem, with one team allowed to use AI tooling and the other not. Then, in the afternoon, we switched roles so that both teams got to try with and without AI assistance across the day. We also worked in pairs or threes. Both problems were very similar: effectively a web based service consisting of a backend API, that used some kind of database, and a frontend UI. If you’re thinking: an online clothing store / IMDB, then you’d be pretty close to the two exercises.

There were some questions about AI in the context of search engines as Google, for instance, puts its AI spin on the result. This was clarified as meaning no direct use of AI tooling for code generation and tooling suggestions. The non-AI teams had to rely on search engines / Stack Overflow, blog posts, docs, etc.

Another question (from yours truly) was about how much of a “hackathon” this really was, or whether we should approach it more like we were building something for real. The answer was that we should try and treat it more like a real project than something discardable like a prototype.

The Non-AI Exercise

I was in Team B, and for the morning session we weren’t allowed to use any AI tooling for our online clothing store. Ironically, trying to disable it to avoid biasing proved non-trivial, while still leaving the traditional IntelliSense and refactoring tools enabled. (Where does one end and the other begin…)

Our group of three had assembled because we all had C# in common. I’ve not done any non-trivial web UI work for twenty years, and it’s also been 5 years since I’ve done any production C# [1] so I wasn’t exactly up on the in-vogue architecture choices. Consequently I left it to the other two to suggest a Blazor based approach, which suited me as I had at least done some work with Razor 8 years ago, so was familiar with the underlying concepts.

The dataset we were given was hosted by Kaggle and came as a ZIP file that included a CSV file and some images. Trying to obtain this without having to sign-up to their service was another small barrier, eventually overcome when I unearthed the CURL instructions which thankfully didn’t require any authentication.

We all agreed to start by creating a walking skeleton that was a simple Blazor app. With some guidance from my colleagues, I used VS Code to knock that up and we were off with a basic web page in the browser.

We then looked at the dataset in a little more detail to see what shape it was in, and there was some discussion about next steps, such as pulling that entire CSV file in. That came off the back of the initial discussion about the model we should use to back the service. Yes, the categories aspect was an interesting modelling problem but I wanted to just to create a simple model with ID, Description, and Price, then hardcode a list of three items, and get that visible on the page to complete our skeleton. That idea was accepted and we got that working quite quickly.

At this point we all got a bit nervous as we hadn’t got any tests in our skeleton. Luckily one of us already had some experience with bUnit (a testing framework for Blazor), so we managed to get a test project and initial test in place relatively easily that checked for one of our hard-coded items being on the page.

This paved the way for starting to add new features in a test-first manner and, with what time we had left of the two hours, we started to work on the test for the product details page which comes from selecting an item on the products page. We had the failing test in place when the metaphorical whistle blew for half-time.

The AI Assisted Exercise

After lunch, where we had a chance to mingle and chat to others about their experiences so far, we tackled the IMDB clone with the addition of AI assistance.

We chose to remain in the same group, while others paired up with different people. We also thought it would be useful to use the same stack as that seemed like a useful comparison, although we did switch laptop and consequently from VS Code to JetBrain’s Rider. Once again there was some fiddling about with settings.

None of us had ever tried creating an entire project from scratch using an AI tool so we threw the README into the chat window prefixed by a sentence that asked it to create the entire solution and projects using Blazor as the stack.

The chat window started spewing out snippets of code which we then started to look at. I think we were all quite surprised about how plausible the code looked. We had some minor concerns about some of its design choices but it all looked pretty sane as a starting point. We hadn’t asked it to generate any tests, preferring instead to let this become an additional change later.

Trying to turn the suggested code in the chat window into an actual solution structure in Rider turned out to be our first time-sink. The UI was a little confusing and the laptop owner hadn’t toggled a setting which we later discovered allowed it to automatically create the files itself. Instead, to make progress in the meantime, we ended up creating the basic solution and project files ourselves, just like before. By the time we found the necessary setting we already had that aspect in place. However, once we enabled it we could at least take the suggestions for the model and Razor pages from the AI output.

This was the point when we discovered what looked plausible wasn’t entirely correct. For example, the code comment suggesting the relative path of the file didn’t match the file it created and it didn’t quite follow the Blazor convention. Hence, we moved a couple of files to follow the expected layout.

However, that said, when we built the code it worked first time and the web page appeared and looked totally sane. One of us remarked immediately that just getting the Bootstrap (UI) set-up right can be a real pain in practice and that alone was a real blessing.

Once we tried to interact with the page though we found it didn’t work. Cue the next time-sink, which was getting the filter feature to actually work. Naturally, we asked the AI tool what was wrong and it very confidently told us what was wrong and how to fix it. But, when I looked closely at the fix I pointed out that it had just re-written the existing code in a more complex manner but left it functionally the same. We went with it anyway but of course it didn’t work.

Another pointless suggestion from the AI tool about how to fix that didn’t inspire confidence at this point, so we resorted to using the debugger to see if the code was even being triggered. It wasn’t, although inspecting the DOM in the browser looked correct. I had raised a question earlier about session state management which didn’t initially sink in until a little later when our Blazor expert realised the page was missing some “interactive server” declaration at the top of the Razor page. Once that was added things started working.

It’s hard to know if the impedance mismatch was because we manually created the project files and the AI generated ones would have been correct for the code it was producing, or if it would still have been wrong.

Sadly, all this friction meant we had little time left to really explore the AI tool in the context of making changes to the existing codebase to add new features or refactor the code. We asked it a few cursory questions to see what it would suggest and it would always reply in a very confident way but would include changes that would revert some of our changes we made during our debugging attempts. I don’t know if that was a result of us not resetting the “context window” because those changes were ones it had originally suggested right at the beginning.

Reflections on the Experience

First up, I don’t know about how the others in my little group felt, but I was woefully underprepared for this day. With only two hours for each exercise we should have been in a position to hit the ground running and I just didn’t have enough experience with configuring or using the AI tool to make the most effective use of the little time available for the exercises.

Related to that, the problem domain wasn’t something I have any real experience of either. While this might have been useful for the “AI as a learning tool” aspect, I’m not sure it helps when trying to compare the two approaches.

With any day like this where you are working with other people you’ve never met before some element of “team dynamics” is going to distort the picture. My preference is to work in small steps, really small steps, and deciding whether to assert that or “go with the flow” is a much bigger side-quest than in an established team where you already know everyone’s favoured approach and politeness levels are well calibrated.

When you only have two hours there is a clear conflict between doing what you really would on a real project versus trying to explore the goals of the exercise. For example, due to my C# being a little rusty I wrote a classic POCO style class for the model instead using the more modern record types in C#, which I was aware of but had never actually used in practice. This prompted a short conversation about our feelings around immutability and anaemic models which is something a real team needs to resolve, but we simply don’t have time for that despite the “not really a hackathon” premise. Likewise, I have opinions on naming, design, etc. which I just have to put to one side this time.

However, I forgot this when writing our second test (in the first exercise) and naturally wrote a test name which opened a discussion about whether we were testing the navigation aspect or the state of the resulting page. Again, I think the approach to testing – the kinds of tests we want to write – is important in a real project. It would have been interesting to see what kinds of tests the AI tool would have generated as I see that being touted as a big use case, but given how poorly tests are commonly written I’m sceptical it has the corpus available to have “good taste”.

As I’ve already said above, I think we were all impressed with how much the tool churned out from the README, but are also aware that online stores and review sites are a well trodden path, with potentially a lot of example code to draw from. Other attendees got a lot further than we did, with one pair (using NodeJS) completing the exercise and having time to make up new features! In the post-exercise discussion one of that pair questioned how it would fair for, say, some embedded software where there are far less examples to train on.

The over confidence in the replies when we were trying to fix the code were scary. I think we were all sceptical and seeing it behave like this did not endear itself to us at all. Plus, trying to revert our code changes and only change the code it should need to when adding new features went against our agreed approach of only changing one thing at a time. Maybe we are being too rigid here and should cede more control in those early moments before we “hit production”?

We only got to work on producing code we had helped to write. It might have been interesting to have swapped at half-time and continue working on one of the other team’s AI generated codebases to see what it might be like to pick-up one of those as that is surely the future situation we’ll be finding ourselves in fairly soon.

Epilogue

I really enjoyed the day and am most grateful for the experience, but I don’t think I contributed that much personally to the experiment because of my current lack of knowledge around using these kinds of tools in practice.

I’ve only used Copilot integrated into VS Code on my laptop but would like to try some others based on the other group’s experiences, as long as they have tight integration into an IDE. I like baby steps and an “always be ready to ship” mentality which I’d feel uncomfortable giving up at this point in time. This is more like enhanced auto-complete based on what I’m already typing directly in the code than trying to write external prompts in a separate window which feels like unnecessary context switching when writing code.

Once I am more used to these new tools I’d like to try the exercise again and go “all in” and see what it chucks out, and then spend some significant time trying to work with that codebase to get a better feel for how much the AI tool wants to change each time. Automating the boilerplate stuff has to be worth exploring, but the question is how much leash to give it?

There is definitely a difference between building a disposable prototype and making changes in a large mature codebase. I think I’ve got more of a taste of the former now (coupled with my own fumblings beforehand) but I’m going to need more time to explore how this fits in with the latter as that’s my day job. At least I feel a lot more informed now than I did a week ago.

 

[1] My client has their own pure, functional, dynamic language which is used for much of the company’s reporting, scripting, and general gluing together of business workflows.

 

A Decade of Lightning Talks with Programming One-Liners

Note: This blog post follows the modern recipe style – a load of unimportant background history before presenting the real content. So, just hit page-down a few times if that’s what you came here for, I won’t be offended.

---

---

---

Just over 10 years ago, fellow ACCU member Jez Higgins replied to another one of my programming one-liners on Twitter with a flippant remark that I probably had enough material now for an entire conference talk…

Lightning Talks

The ACCU conference generally hosts 90 minute talks. Of course, there was no way anybody would sit through that length of session, even if I had the relevant amount of “quality” programming puns, and if the conference committee were crazy enough to accept it as a serious proposal.

However, another popular feature of the ACCU conference were the lightning talks, which were only 5 minutes long and held at the end of each day. The barrier to entry there was considerably lower too as they could cover virtually any topic. Also, I’d already given a few of these in the past, all with programming related humour being the general theme, e.g.

2010: Recycle Bin 101 – a Room 101 affair featuring some of Microsoft’s tools which refused to die gracefully back then, such as Visual C++ 6, Visual SourceSafe, and IE 6.

2012: Not Only, But Also – a riff on the #NOSQL, and #NOESTIMATES movements featuring my own variation: #NOSHIT.

2014: The Art of Code – a selection of code snippets framed as art, along with humorous titles. (This was inspired by Chopin’s Waterloo at the Pompidou Centre in Paris.)

2014: Requiem for Windows XP – a recital of Rutger Hauer’s final words in Bladerunner followed by the shutdown screen for Windows XP to mark its demise by Microsoft as it had finally reached its official end of life.

An Actual Comedy Routine

With Jez’s tweet in mind I decided that for 2015 I would go for a more formal stand-up comedy routine and read out what I hoped were some of my “best” and most appropriate puns for the ACCU conference demographic, which are typically highly technical programmers, largely from the C++ community, and mostly of a certain “vintage”. With the rise of the agile movement in full swing at the time, “The Daily Stand-Up” seemed like the perfect play on words to use for the title. The set of 30 or so puns were delivered in a deadpan style to an audience of 300+ attendees, and the reaction was such that I considered it a success and definitely worthy of further attempts in the future.

Over subsequent years I “performed” similar routines at Agile of the Beach, the Equal Experts Christmas Party, Agile in the City Birmingham, NorDevCon, my Spektrix leaving do, and most recently, the D language conference. While they were mostly only an occasional gig or two, the lightning talks at the ACCU conference became a more regular fixture (aside from the Covid years when it wasn’t run / was purely online).

10 Years Later

And so here I am, a decade later, at the ACCU 2025 conference with another mixture of programming puns which features some old favourites, more recent material, and even some as-yet-unpublished (i.e. not tweeted, yet) material. So, if you missed out on the “live performance”, or are simply a glutton for punishment, then you can relive the experience (groans and winces entirely optional) by reading these to yourself. Enjoy!?

“The background AI in my text editor just asked me: are they paying you to write this crap? I thought that was a bit harsh, but then remembered I’d enabled dark mode.”

“I’ve started creating a typeface where each glyph is a different page from Wikipedia. It’s going to be the font of all knowledge.”

“If Bitcoin is value realized as ones and zeroes, does that effectively make it Gold Boolean?”

“Have you noticed that you never get a simple yes/no answer from people in the Tri-State area?”

“If I join the frontend channel, backend channel, database channel, and operations channel, does that make me a Full Slack Developer?”

“Now I do more working from home I like to get my kids to help me with my coding. I call it au-pair programming.”

“The trouble with pair programming in C++ is that your codebase becomes littered with .first and .second.”

“Our company is a big fan of the Pimpl Idiom. They like to hire spotty teenagers that vibe code C++.”

“I’ve always felt algorithms that round down to the nearest integer are inherently floored.”

“When first learning C++ I was given advice that I should ‘do as the ints do’. So now I make sure my code behaves slightly differently on different platforms.”

“Is a Polyglote a programmer that boasts about how many programming languages they know?”

“I once asked the Enterprise Architect if there were any non-function requirements? He said: yes, you can’t use Lisp, Haskell, F#, Elm, …”

“He clearly favoured an object oriented approach –anything we wanted to try that was different, he’d simply object.”

“I wonder if Google have postponed the next major version of their popular programming language after Dijkstra asserted that Go 2 is considered harmful?”

“I was really excited to discover Tony Hoare and Dennis Ritchie were giving a talk about two of their most influential contributions to Computer Science, but quickly became disappointed when my tickets arrived and said ‘Null & Void’.”

“People who are PRINCE certified just want to run projects like it’s 1999.”

“I recently tried to use an LLM to write some code to produce a digital certificate, but it just made a hash of it.”

“The infosec team asked us if we regularly rotate our keys. I told them: yes, we often do that by passing them through ROT-13.”

“I once debated with Alonzo, the creator of Lambda Calculous, about whether lambdas and closures were the same thing. He said ‘No!’, and that’s when I discovered there was a clear separation between Church and state.”

“We’re really regretting hiring this guy Sisyphus as a DBA, everything he does just gets rolled back!”

“In the battle between relational and document oriented databases, I think it was when SQL introduced the PIVOT function that the tables were turned.”

“We recently recruited a prompt engineer. It’s great that he’s in the office every day at 9am sharp, but then he spends the rest of the day tweaking the PS1 variable in his Bash config.”

“Have you ever noticed that off-by-one errors know no bounds?”

“The Wildlife Trust has deemed our codebase a conservation area on account of the number of bugs.”

“Our product owner asked why our acceptance tests only cover the happy path. I told him: it’s because they’re rose-tinted specs.”

“The Marketing department said that as a company we needed to be more disruptive, so I dropped the production database and deleted all the source code.”

“The other day I passed a chap typing fast and furiously to try and quit his text editor. I think he was called Vim Diesel.”

“I reckon my colleagues only find my jokes amusing when I’m in the office. At least, I think that’s what they meant when they said: you’re not remotely funny.”

How Do I Test This?

I’ve no idea where I first came across the following premise for when starting a new code change:

“First ask yourself, how do I test this?”

I had assumed it was Steve Maguire’s Writing Solid Code or Debugging the Development Process as these are two of the earliest books I read as a professional developer. He has some excellent advice in “step through your code in the debugger” and “how could I have prevented / automatically detected this”. But it wasn’t either of those. No matter, I just like to remember my sources of inspiration.

Naturally, one shouldn’t take this advice too literally as there are other questions that really need answering first, such as whether this feature is even needed. The point is that once writing code is part of the solution, then thinking about how you’re going to test it should be right up there on your list. One behaviour that is common in those just starting out in their programmer journey is the desire to jump right in and start writing code without any thought for how they are going to answer the eventual question “does it work?”. And, by extension, “does it solve the customer’s problem?”

For a bug-fix, that might seem obvious, if the bug is easy to reproduce, though there is still the follow-up question of “have I broken anything else by accident?” Discovering up-front what you might break can have a sobering impact on your approach.

For new features the question can be much harder to answer, especially if the change is buried deep inside some complex system. It’s all very well being able to change some code and assume it works because the dry-run in your head suggests it will, but it’s another thing to be able to show that it works by running it in an observable way. This is doubly true when you factor in that famous quote from Edsger Dijkstra:

“Testing can prove the presence of bugs, but not their absence!”

Design for Testability

What both of these quotes try and convey is that testing requires a change in mindset, it’s not just some activity you throw in when it’s “code complete”. I’d hope in this modern era of software development that we’d all recognise the folly of believing that you can just bash out some code in a non-trivial codebase and assume it’ll work first time. Or even that you’ll obviously be able to see all the potential edge cases up front. (The average developer only spends 2-3 years in a role so we’re constantly putting ourselves back in the position of “new joiner”.)

To be clear, I’m not specifically talking about TDD (Test Driven Development) here, at least, not in the formal unit test / Red-Green-Refactor sense. I’m talking about it in a more general sense of how thinking about how you’re going to test something will affect how you approach the task at hand.

Plenty of what I do on a day-to-day basis does not have any formal automated tests produced as a by-product of it. For example, I am often heavily involved in the build, deployment, and support side of software delivery as much as writing code for the main product. While the latter will have a barrage of automated tests to help validate every change, the former often relies on an amount of manual testing. That’s because the cost benefit of introducing automated testing to a bunch of ad-hoc Bash glue scripts which rarely change is typically very high in comparison to the cost of testing them manually.

However, what makes the cost of testing them manually much lower is not the inherent nature of them, but the mindset used in how they were written – they were designed with testability in mind. What that means in practice is that instead of writing only the bare minimum to solve the problem at hand, some additional effort is made to allow it to be tested safely and quickly, both now and when future changes are made. It’s just another trade-off we have to be cognisant of and weigh up whether it’s worth the extra effort up-front – is being easier to test worth it? If so, how much easier is enough? (There is always an XKCD, and Is it Worth the Time along with Automation are well worth the time revisiting every now and then to remind yourself of the potential costs and pay-off, and delusions of grandeur.)

Risk / Confidence

When I say “easier” you shouldn’t just assume I mean “quicker” either. Testing is about risk, and developing confidence. The goal is to deliver the highest level of confidence with the lowest amount of effort. Printing out a log message saying that you would have deleted a particular file gives you confidence that you’d have deleted the correct file, but can’t point out that the entire process aborted because another process typically has it locked and now you might need to add logic to cater for retries and partial success. The principle of YAGNI (You ‘Aint Gonna Need It) is a strong force for reining in purely speculative requirements but can also be weaponised as an excuse for why anything outside the happy path was ignored.

My point here is that you consciously consider how you’re going to gain confidence and weigh up the options. Maybe when staring at that log message you’ll ask yourself “what would happen if that file is locked?” and consequently you go and adjust the code to allow you to target a different folder where you can create a locked file and test your hypothesis. Is that a temporary hack or do you introduce a command line switch / config setting [1]? If you’re going to keep hacking the code every time you test new changes then you run a higher risk of committing that hack by accident and breaking production [2].

Scope

We could very easily expand that original phrase “test this” to the more specific “test what happens when it fails”.  Only testing the happy paths is such a common malaise, although I suspect it is largely borne out of naivety rather than short-sightedness. The cost of debugging code that fails in a horrible way is huge, and frequently occurs as the result of a production incident when time is of the essence. Taking the time up-front to think about failure scenarios and test how they will manifest will really help reduce stress levels when the inevitable happens. (Stack Traces are a crutch we lean on far too often when a good error message could provide better value.)

Another tweak would be to replace the “I” with “we” in the original proposition. The person who finds the code difficult to test in future might still be you, or it might be a colleague. I don’t particularly like that analogy of writing code as if some psychopath will hunt you down later, and prefer instead to try and promote “paying it forward” as simply the right thing to do. The ultimate outcome is a healthier team, and codebase, and a faster, more sustainable delivery pace. Doing it right actually benefits the business in the long run.

Always be Testing

The question “how do I test this?” doesn’t necessarily stop being asked once the code has been delivered either. As one of my earlier posts (Validate in Production) shows, Dijkstra was right, and we need to be more pessimistic about our ability to get things right, first time, every time. Remember though there is a balance here. If you take Dijkstra at his word then you’ll never deliver anything as you’ll spend your entire time trying prove the absence of any bugs. (Formal Methods might be the exception here but I have no experience of that, or know anybody that has.)

It’s also only a short hop from that question to a closely related one – “how do I support this?” Ideally any separate support or operations team are already recognised as stakeholders in the system and their interests are explicitly factored in [3].

Taking this yet another step further leads you into the whole world of observability and ultimately around testing the business hypothesis of whether building that new feature delivered the benefits expected. Having that mindset where you can work backwards from “how to test the business impact” has yet more benefits that can be reaped.

Remember Overthinking is not Overengineering.

 

[1] See Testing Drives the Need for Flexible Configuration for further thoughts on that approach.

[2] This is one of the many things in my mental Commit Checklist which has come from making these mistakes myself in the past.

[3] Both Support-Friendly Tooling and the much older From Test Harness To Support Tool look at this angle.

Unravelling Conflict

An interesting question came up at a meet-up I recently attended, where they were discussing approaches to coaching and mentoring. The question was not overly interesting by itself, but the roller-coaster of a journey as we got to the essence of the real problem, was. The initial framing of the question by one of the attendees was this: 

“How do I tell someone to optimise their code?”

The wording here immediately made me feel a little uncomfortable because I generally dislike the notion of telling anybody what they should do or how they should behave. This was a session about coaching and mentoring and I felt that before we could answer the question we first needed to talk about the question itself, and the language used to express it. It helps immensely if you don’t go into the process assuming you are inherently right. But maybe it was me that misread the tone of the question?

One of the most useful guidelines for navigating life (not just the work environment), I first heard from Jeffrey Snover – Chief Architect of PowerShell. He often used to tweet this saying:

“When faced with conflict, respond with curiosity.”

In this instance I’m the one facing conflict. The tone of question is at odds with how I feel people should behave towards others. In the past I have occasionally had the misfortune to work with some people that had very different ideas about collaboration (see Fallibility) and so I also know that sometimes the answer is simply to walk away.

However, the attendee was very likely not a native English speaker, and so there was every possibility the choice of words was simply unfortunate. They also seemed like a nice person from previous conversations I’d had, so this really felt like there was something much deeper going on that was worth pursuing.

Unfortunately, the situation got slightly worse, before it got better. Some of the other attendees immediately started to challenge the assumption that anything might even need optimising. After all, the general wisdom on this topic is “don’t do it”. And, for experts only, the second rule is “don’t do it, yet”. I was happy to see where this line of questioning went and even added my own typical questions, such as, “have you measured it?” and “what business value would rewriting it bring?” Ultimately this lead to the even more uncomfortable statement:

“I just don’t like how it’s written.”

I’ve worked on a few codebases where there is a lot of noise in the history because people decided to arbitrarily rewrite code simply because it didn’t fit their personal preferences. In the worst case code flip-flops between two styles as any chance to touch the module becomes an opportunity to rewrite it, yet again. Doing this just creates conflict within the team and is ultimately very wasteful. In contrast, genuine refactoring, that brings to code forward to adhere to the team’s current preferred principles and practices does have some merit. (See Sympathetic Change for more thoughts on this.)

Anyway, this is going badly as I’m finding it harder to empathise with the speaker and my opinion of them shouldn’t really cloud my decision to help them. After all, they’ve come to this meet-up because they want to learn, like the rest of us, and we should fully embrace that.

And then the magic starts to happen.

Someone asks what it is about the code that they dislike, and we find out that the code was written using some very poor choices. The speaker continues to describe how there are more modern idioms which are inherently more performant and in-keeping with how code tends to be written in that language nowadays. Hence, this isn’t so much about personal preference as de-facto standards.

Now we’re on a roll.

Further questioning reveals that this code was written by a relative newcomer, i.e. a more junior programmer. The original question was badly phrased and the real question was about how to approach a more junior member of the team and mentor them on how to write better code – code which will perform better out-of-the-box, and also fits in with the industry norms. (Writing incredibly naively is known as pessimisation – the opposite of optimisation – and we shouldn’t make poor design choices purely to avoid being seen to be prematurely optimising.)

Hence, what had seemed at first like a somewhat selfish question was actually altruistic in nature, and entirely within the community spirit the meet-up wants to promote.

In retrospect I realised this was effectively an application of the classic Five Whys technique often used during incidents and post mortems to get to the “root” cause [1]. It’s easier to apply that technique to an incident because it’s less personal, whereas the continued questioning of a single person can feel more like an interrogation, so you need to be very careful not to appear to be attacking the person. Responding with curiosity is key to enabling that process as it allows you to create a warmer, safer space and therefore allow the other person to open up and become more reflective.

 

[1] There are problems with the notion of there ever being a single cause of a issue, hence the use of quotes. How Complex Systems Fail is mandatory reading on this topic.

 

Using CoPilot-Like Tools is Not Pairing

Ever since the rise of LLM based tools like ChatGPT and CoPilot there have been quips made about how great it makes as a pairing partner. The joke is because these tools are passive and won’t trash your code, or disagree with your approach, or smell of garlic.

I get the joke.

The problem for me is that these jokes are suggesting that using tools like CoPilot are akin to pairing, but they’re not. And, as a consequence, they are continuing to perpetuate misconceptions about what pairing is really about. This topic came up one Tuesday at a Cambridge Software Crafters meet-up when we were discussing mentoring & coaching and how to adopt pairing in companies that are hostile towards it, likely for these very reasons.

(If you’re after a longer read on this subject I can offer you my 2018 C Vu article To Mob, Pair, or Fly Solo, or if you’re a video kind of person then you might want to watch the talky version I gave at the ACCU 2024 Conference, also titled Mob, Pair, or Fly Solo.)

Tools like CoPilot are there to help you with the “mechanics” of writing code [1]. They complement the traditional tools like compilers and linters which help avoid a bunch of annoying problems that come with creating software as part of a solution. Your pairing partner may help you avoid these mistakes too, especially if you’re new to the language or tool, but that’s not primarily what they are there for.

Their input should focus on helping you to solve the actual problem. That might be at a design level, such as how best to represent the solution using common idioms or patterns, because they can see where it might be heading. They can also help come up with potential problems or edge cases that need to be factored in, and help decide how and when that might happen in the evolution of the feature you’re working on.

I’ve never paired in the strict navigator-driver sense, it’s always tended to be far more free-flowing, likewise when ensemble programming too. The keyboard tends to go back-and-forth with “type, don’t talk” being the trigger to switch, when it’s not blindingly obvious what my partner(s) are trying to convey.

Hence, in a TDD style loop the other half of the pair might already be thinking about the next test or where the design might be going for the refactoring step after the test passes. Sometimes one half is on a roll and it’s better to maintain momentum, other times both minds need to focus on the exact problem at hand, such as when the test fails unexpectedly. (If you both missed it then that is a smell worth exploring because something is clearly amiss somewhere.)

With modern IDEs and syntax highlighting the typist knows when there is a typo, they don’t need to be told by their partner. However the way they drive the IDE and other tools can often be a point of discussion as there are often many shortcuts and extensions that provide easy wins. High-end tooling often has an ability to try and point out other features you might find useful, and ChatGPT like tools could probably summarise related features in popular tools if asked.

Humans often remember things better when the correction happens at the point of triggering, and what your partner can see that your tooling can’t, is whether or not it’s an appropriate moment to make  a suggestion. There is a time and a place to share knowledge and part of the art of pairing is to read the mood and only interject if it’s likely to be beneficial now, or note it for a retrospective conversation later. It may even resolve itself automatically, later.

Sharing knowledge takes practice. For instance, sharing the right level of detail or techniques and idioms, to avoid overwhelming your partner while they are trying to focus on the problem at hand. The new generation of coding tools might be well versed in programming language details and common patterns but they don’t know anything about your organisation, team, or design principles. They might help with the small picture but they can’t help with the medium and bigger pictures, and that is the bit that makes software delivery hard. That is also the area where your partner should be adding the most value to the exercise.

Programming is so much more than just writing code, and programming in pairs and groups is about active participation to triangulate on a sustainable solution far quicker than working with passive tools and feedback.

 

[1] Maybe one day they’ll be able to consume your entire codebase, along with the VCS history, so that they can tell what’s old and what’s new and make better design suggestions too based on evolving trends. We’ll see…

 

Crafters Meetup: Architecture Kata

Last time I wrote about the monthly Cambridge Software Crafters Meetup which I started attending almost a year ago. In that post I briefly mentioned the Architecture Kata session which I found particularly interesting as I’ve never done anything like that before.

I’ve been aware of katas for a long time, but they’ve tended to focus on solving simple problems in code, usually with a test-first bias, but this is the first time I’ve seen it applied to other aspects of software delivery.

Session Format

The attendees divided into groups of four or so people and were given a problem statement to design a system. We then spent an hour or so discussing it before presenting our design and approach to the other teams at the end. One of the organisers acted as a sort of product owner who we could ask for further background or clarification.

The problem we were given revolved around the overnight charging of a fleet of electric vehicles. The statement contained a fair bit of information about the charging rates and prices, along with other bits about the scale and monitoring.

Each team had a large sheet of white paper and a pack of sticky notes and pens to use.

Big Picture or Small Picture?

One of the things I’m trying to get out of these meetup sessions is to sit back and see how other people approach problems. Having 30 years of experience means you have plenty of biases to overcome, but at the same time I’m not going to sit there and watch my team spin its wheels stuck in Analysis Paralysis or get lost down in the weeds. Hence I was prepared to try and nudge when I felt it was appropriate as we had limited time, but mostly I wanted to avoid appearing to take control of the group and drive the direction.

To me, architecture is generally about big picture stuff. In this particular case: what is a suitable overall shape for the solution, such as computation, storage, network connections, etc? For example, the problem statement suggested that efficient charging was desirable from a cost perspective, which implied to us some kind of scheduling component which could drive the remote charging stations. The potential for failures (of vehicles, charging stations, etc.) also implied to us that this scheduling must be dynamic. However, that’s as far as our team went, aside from noting the need to store tariffs somewhere and perhaps historical data if charging was not effectively uniform to allow the system to be adaptive.

One other team also looked at the problem at roughly the same “scale” we did, although they went as far picking specific technology choices as they had a firmer idea of exactly what services they were building and how they wanted to connect them.

Somewhat surprising to me, two other teams focused deeply on the algorithmic element of how to efficiently charge the vehicles. Their presentation was all about what we had simply called “the scheduler”, and then ignored all the other stuff that sits around it. In my mind these were implementation details rather than software architecture.

Clearly the point of the session was to learn, so there were no right or wrong answers, but what in retrospect I’d be interested in knowing now is what the other teams think software architecture is about. The other thing that occurred to me is that maybe I drove the direction of our team’s solution more than I thought, despite trying hard not to.

Actors

What also surprised me, and maybe this was one of the reasons why we ended up focusing on the big picture, was that our team was the only one which had any actors (i.e. users) in it. This was an early suggestion of mine as we were talking about the problem statement but nobody seemed to be looking at it from the user’s point of view. After all, software development is about solving people’s problems and it’s not always obvious who those people are at first.

There weren’t many users mentioned in the problem statement per-se, but I was trying to use that question to open up a wider discussion about who the stakeholders really are, and how those might affect the shape of the system. And not just the shape, but also the evolution of it as it grows from a walking skeleton to a fully-fledged system – the journey is often far more important than the destination, which we know will change over time.

Design Surface

One thing I did do early on was to draw my actor on a post-it note and stick that on the sheet of paper after I noticed someone was about to draw directly onto the sheet. The other teams sketched elsewhere and then drew up a final diagram and notes on their large sheet. Our final presentation was essentially a sheet covered in sticky notes.

I’ll be honest, I love sticky notes :o), but I think one of the lessons of the architecture kata is that it should be fluid, and in the early days a system can be incredibly fluid, aka emergent design, therefore the temporary nature of sticky notes helps to explore and convey that. Once we had positioned some of the key actors (which weren’t necessarily people) at the edges we could then break down and refine what happens in the middle without having to re-draw from scratch or squeeze things in. (You don’t need an electronic drawing board at this level could have been another lesson.)

One of our team got some scissors and started cutting up the problem statement into little sections and attached them to the relevant part of the diagram. I really liked this as it gave a physical manifestation to the decomposition of the problem. It also helped us keep track of what we had covered and what we needed more detail on.

Epilogue

I really enjoyed this session, mostly because greenfield architecture is something which we rarely get to do as software developers. The vast majority of what we do is maintenance on existing systems and, even when adding new services, we are still designing within a space that is largely already well-defined. In my 30+ year career I’ve only been in this greenfield situation for real maybe three or four times.

My background is largely in financial systems so the problem setting was new to me, although once you break it down you find that dashboards, scheduling, etc. are all smaller problems I’ve met before so have other experiences I could apply to this particular scenario.

One consequence of the problem setting – the physical nature of chargers and electric vehicles – and my lack of knowledge of this technology, meant that these were big risks to me; so my general thoughts were around how to address those and whether that might have an impact on the architecture. Ever since reading Waltzing with Bears: Managing Risk on Software Projects I have tried to factor the identification and mitigation of risks early on in my design thinking.

I really forward to another architecture kata session. One point of katas is to solve the same problem multiple times but by using different approaches. I’d love to do that but I suspect the desire of the meetup group would be to do something new each time.

Our team definitely didn’t make enough use of the product owner, which may have been because we weren’t sure whether that was really the role being played, or they were just there to provide some basic clarity. In the exercise I described in Choosing a Supplier: The Hackathon we had a literal product owner shared amongst teams so that was more like what I was expecting.

It would also be nice if we could find a little time to discuss what we even think “software architecture” means these days, and also reflect more on how we approach it as this session raised some interesting questions for me around the very nature of this level of design.

Cambridge Software Crafters Meetup

Despite living pretty close to Cambridge I’ve never actually worked there. For the past 25 years I’ve commuted to London because it was easier and faster than getting to Cambridge [1]. The daily rates for contract programmers in the finance industry might also have had an influence too :o).

Anyway, one consequence of working in London meant that I had access to all the London meetups in the evening. But I also had a wife and four young children, so wanted to keep the balance right and consequently I largely limited myself to the ACCU London and eXtreme Tuesday Club (XTC) events, only dabbling in others if there was a particularly interesting talk / topic.

Then the pandemic happened, and like many others I found myself working remotely full-time. I had worked from home before on previous contracts, and in some cases that was to facilitate being able to attend a meetup in Cambridge, most notably around .Net and Go, but I only attended a handful or so. Mark Dalgarno (via Software Acumen – organiser of many Agile related conferences, such as Agile Cambridge) was in Cambridge back then too, and through Software East also arranged various enjoyable gatherings.

Eventually some semblance of normality resumed but I still found myself working remotely and now Cambridge was my nearest point of social contact (IT wise) in the evenings. The pandemic meant that meetups had to stop for some considerable time (years!) and with people nervous to venture back out again the choice was naturally limited. Platform focused meetups like Cambridge .Net reappeared in-person sooner, as I guess they had a captive audience, but I wanted something that covered the wider topic of programming at large, not just one specific language / stack. (There was an ACCU Cambridge meetup once-upon-a-time but sadly that folded a long time ago.)

Tangentially, my eldest son had also recently entered a career in programming and he asked me about any local meetups I knew of as he was based around the same area. It didn’t take long to discover that the Cambridge Software Crafters group on Meetup was holding its first in-person gathering since November 2019 (a hiatus of four years). The topic was learning TDD and it was being held in the office where my son was working at the time so I suggested he had no excuse not to attend either :o).

I met some old faces at that event and, more importantly, some new faces too. (I even bumped into a young chap who lived a few houses down my street. I didn’t know he’d also started a career as a programmer.) A diversity of ages, experience, industries, culture, etc. is an important requirement for me in a more general craft-oriented meetup because “one size never fits all” and I like to be aware of what forces and constraints drive other people’s work.

For that first meetup we did a leap-year kata in pairs. I chose to work with someone young who wanted to use TypeScript as I’d never used that before. Having practiced TDD for almost 20 years I clearly wasn’t there as a beginner but I felt my experience might be useful and I always learn something new when pairing with people. (After pairing / mobbing heavily for the 6 years prior to my current contract I realised how much I had missed the sociability of ensemble programming.)

Sadly I missed the next few events but my son kept me informed on what they had got up to and it continued to sound interesting, with practicing TDD clearly a popular topic. I eventually attended the Architecture Kata, which I found interesting and intend to write up some particular observations in a separate post. Then we did an evening of lightning talks (10 mins) which is something I’ve always enjoyed at the ACCU and Python conferences as you always get a pretty diverse set of topics. This was no exception. (It also gave me an opportunity to resurrect a humorous short talk I did at ACCU 2014 – The Art of Code.)

The most recent meetup was another TDD kata, this time around the game of Noughts & Crosses. I ended up pairing with someone that was only just getting into programming and was attending the meetup to learn more about the craft as he had a background in construction. In the end we hardly did any programming per-se and I mostly walked him through TDD but more as a backdrop on talking about what sustainable delivery really entails. It wasn’t at all what I was expecting to do that evening but I was glad I could share some knowledge & experience, and it meant that everybody then got a suitable partner they could work with.

While the sociability factor is high on my list it would be remiss of me not to mention that the event has sponsorship from Codurance which means that the evening comes with free pizza and drinks. The venue, The Bradfield Centre, is a nice modern building and has an auditorium which was used for the lightning talks. Most events take place on the open plan ground floor area as we either pair or work in small groups and then share with everyone what we’ve done / learned.

I really hope this meetup continues to gain traction as I think we often focus too much on specific technologies and not enough on the craft as a whole. There are clearly a few regulars but also new faces each time which provides a nice mix and everyone has been really friendly. It also gives me an opportunity to “talk shop” with my son as that never goes down well at family gatherings :o).

 

[1] Things are much improved with the guided buses and upgraded A14 but what I wrote in “Missing the Daily Commute by Train” still holds as far as using the train goes.