Concrete and practical advice for dealing with a widely shared issue in large software environments with ISO and other certifications to cater to. Thanks for sharing an article about modern AI assisted engineering without the slightest hint of catastrophising; it’s a breath of fresh air!
This is a nice frame for thinking about the divergence in attitudes towards coding with AI assistance in different environments; thanks for sharing. As someone who’s spent decades in brown field environments, my starting assumption when starting in a new environment or project is that the code will be anything but ideal, with historical layers of sediment from different eras of trends evident. I expect it to be poorly and sparsely tested, inconsistently patterned, and unpredictable to change. Most old coding curmudgeons I speak to share these expectations. What I don’t understand is why more people don’t see AI as the greatest assistant in establishing your understanding of an unfamiliar system, and why they think the code AI is writing is going to be worse than all the code we have to deal with that was written by humans in the decades up to now. Because AI is also the best assistant I’ve ever had in both establishing understanding and making leveraged changes to such code bases. But you can’t just prompt your way there expecting code changes from every prompt; you have to pants your way to a minimum viable understanding, then plot your way to well planned changes using intermediate steps such as those encouraged by the Matt Pocock skills.
This article takes me back to the Martin Fowler book many years ago that for some reason was far less popular than most of his were at the time on DSLs. Reading the DDD book around the same time plus Novak’s Learning How to Learn was the conglomerate inspiration for creating the simplest possible DSL for capturing any domain’s ubiquitous language. it’s simply concepts connected by predicates, one sentence per line in a plain text file, and while it works well to keep LLMs conceptually aligned, they do get big quickly and need a tool to visualise and edit them so I built one: https://thinkingtools.software/concepticon/
I think this class of tool is solving a problem that needs solving, and an “amnesia tax” is a great way to frame it. I had PMB on my list to evaluate for this purpose and I’ll add yours to my list. Are you able to compare and contrast it with PMB? https://news.ycombinator.com/item?id=48631169
Thanks for sharing an honest and balanced take after going down a path that many people would be wondering about. In the last six months I’ve built a product from scratch that I’m in the process of commercialising and I have also found a renewed appreciation of all the things that aren’t the coding, the latest being the hoops I need to jump through to get satisfractory app signing going for the Windows version of my product as a Mac-based developer. Your issues with customers not liking what you give them after a wrong assumption gets baked in rings true to me, and I’ve gained a lot more confidence using some of the spec-driven development tools and as of this week, Matt Pocock’s new /wayfinder skill to make sure the path through the fog is clear enough to proceed before starting. highly recommended.
I've always loved and used flashcards, and relate to this article. I have built my own learning tool based on concept maps, which was inspired by Joseph D. Novak's book "Learning How to Learn". Concept maps are focused on how large networks of concepts connect to each other forming large graph spaces. One of the key problems after creating or encountering a very large map, though, is how to approach them in chunks. So I built a flashcard-inspired feature that chooses a concept at random and focuses your view on just that concept and its directly connected concepts. It's effectively a conceptual flashcard generator. You can check it out here:
https://thinkingtools.software/concepticon
It's a (sorry, Mac only at the moment) desktop app that uses plain text files in a format you can easily edit yourself in any text editor.
It has always mattered for a team to be on the same page about what words mean in your domain. In 2026 it matters even more, because your newest team members are AI coding agents — and they need that agreement in an even more accessible form than humans do.
Concepticon stores everything as plain-text propositions you can read, edit, and diff. One set of propositions; as many diagrams as you need. The propositions are dense enough to serve a human reading the visualisation, and dense enough to serve an AI agent reading the text. No cloud, no proprietary format, no lock-in. Your domain knowledge is yours.
Since the late 2000s, in collaboration with friends like Simon Harris and Sebastian von Conrad, I've built various versions of this — culminating in conceptmaps.io, which I shut down in mid-2024 after the Twitter sign-in broke and I no longer had the technical skill or time to fix it. AI coding tools have now made it possible to rebuild it better than ever in my spare time.
The free Reader is for viewing, navigating and sharing concept maps — no purchase, no account, no email. Concepticon Pro ($50, once) is for the people who make them: in-app proposition editing, AI map generation, document and URL import. This first release is macOS on Apple Silicon; Windows and Linux to follow.
This is an easy to understand and apply approach, thanks for sharing. To go deeper, I might apply the thinking processes from the Theory of Constraints. As Goldratt used to say, “there’s an assumption under every arrow” in the diagrams you generate to map your logic, and uncovering them using the defined categories of objections in that process also help with facilitation among opinionated people who don’t initially see eye to eye.
I think everyone is right on this question :) I have certainly worked in places where there were enormous code bases written in dead languages going back to the 1970s and I was part of the collective belief that nothing could be done about it and our job in the 2000s was to put lipstick on the pig by burying it behind a web portal, if you remember that short-lived fad. In that kind of environment I would have _loved_ an exact port to a modern tech stack from which we could begin the very slow and careful evolution. Speaking to people who work there, AI has indeed changed at least their perception of what is possible and they have traction on porting it that was considered impossible for decades. Whether it works for some definition of “works” remains to be seen, but it might be self-fulfilling because the belief that it can be done will mean it is tried more often and some of those tries will probably succeed.
With that said, I’ve also recently done a rewrite in a completely different sense, taking what used to be a web app and rebuilding from the ground up as a desktop app instead. Having the original code base for core concept reference, but rethinking the whole UI more than a decade on was IMHO a much better approach in that case.
I don't _think_ so, but my understanding of BMAD is second hand. I gather it provides multiple types of agents who play different roles but that those roles are predefined. The frame work I am using lets you "hire" people of any kind of role, based on a job description, a personality, an autonomy ladder and access to tools that you define, and it operates in a session-based mode. It's a bit like having 1 on 1's with different experts and it has helped me immensely with my knowledge and skill gaps on the commercial side.
this issue resonates for me very much, as a one person startup experimenter in 2026 like a whole bunch of other people. I have a day job with lots of human interaction, but I've been trying to find the edges of what's possible as a single person on my fun side project, and I still haven't found them. I built a product in 3 months, and then wondered if I could commercialise it in another three, which I've now done. The unlock for me was an agent framework a friend made that allows me to "hire" collaborators that don't just take orders, but push back, do research and challenge my thinking. I've made a virtual head of product, a virtual head of marketing, a chhief of staff, etc, and can only say I've been blown away by what they've added to my thinking and my capabilities. They totally reframed my approach to building my web site, choosing a commercial back end, and commencing the marketing journey. It's not magic, but it encodes a management philosophy that resonates with me and I still have no idea where the limits are with what I can do alone...
https://github.com/normannoble/agent-framework
I'm a huge fan of looking back to sources outside our own field for ideas on how to cope with the changes within it. An influential one on me comes from education, too, "Learning How to Learn" by Joseph D. Novak. The main take away is that our brains work associatively, and you can both extract and share mental models into a visual form called a concept map. I've been experimenting with capturing them as plain text files that both humans and agents can read as a very dense form of shared memory. When trying to build understanding, the idea is you start by extracting the associations people already hold, and then build bridges one proposition at a time to the new concepts.
This makes sense to me, and reminds me of the more familiar variation in rates of chage in building business systems. I think I first came across this in the writing of Ronald G. Ross in relation to his business rules approach. At the bottom layer, changing at the slowest pace, are the core business terms/ concepts, above which sit business rules, and above those business processes, and above those user interactions, which change the fastest of all. The current trend we see to provide agentic end user access to business processes formerly intermeidated by web pages is an example of the friction from above placing demands on the levels below to adapt. If the business process layer isn't set up well to serve completely new kinds of user interatction, then it will come under pressure to be refactored in order to do so from the faster pace of change at the higher layer.
Yes it's a product of mine, but it's actually a recent rebuild in desktop app form of an app I built as a web app well over a decade ago. And yes, I used it to get hmans on the same page at some scale, which you can hear me describe in talk I did at the time like this one:
https://www.infoq.com/presentations/concept-map/
I haven't done formal evals in this AI era about how having a concept map alongside the codebase changes what the agents do, but informally I know mine read them because I tell them to and they seem to take it into account, and I seem to have less trouble than the average geek I talk to keeping my code base adequately coherent. It's just plain text context, after all so there's no magic, just very concise and accurate context.
This article describes one of the classic polarities in our industry, the tension between the short term and long term. There are many others, such as the tension between autonomy and alignment in organisations with lots of teams. The article did not mention polarity management, but it would help it if it did IMHO. The essence of polarity management is that the problems described can never be solved once and for all, because it's not a problem that you can fix, like you can fix, say, the damage to your car after an accident. The short term and long term tension will be ever present, and the key to managing them is to acknowledge that both have positive attributes, but both, when taken too far, have negative ones. The best advice in these kinds of situations is to work out in which direction you may have leaned too far, and lean back the other way. The key is not to dogmatically lean back too far the other way because you've treated it as a problem that can be "fixed". It can only ever be managed, and each fix creates the next imbalance.
this looks very interesting to me and I plan to try it. I was looking at the web site (pmbai.dev) earlier today and it looked very thorough but when I went to show someone this afternoon it no longer loads in my browser, saying it can't establish a secure connection. Verified by clicking the website link on the repo page on github so I am pretty sure I have the url right...
Depending on your definition of context, my attempt to solve just the foundational context of what words mean in a given domain, and sharing it cleanly, quickly and reliably with both new people in the team and AI agents, is the humble concept map stored in plain text.
Talk here: https://www.infoq.com/presentations/concept-map/
Product here: https://thinkingtools.software/concepticon
Ha — you actually did it. Genuinely made my day; you've now read it more recently than I have. And I think you've landed where I did: the concept-mapping part is the keeper and the rest I have not put to practical use.
I'm intrigued by this idea, and plan to test it to build a new product that is a sibling of an existing one, but with a different targeted purpose. I believe it's verifiable enough to try a goal oriented approach but I'm slightly nervous about this all just being a way to get us to burn the next order of magnitude in tokens!
I'm about to try Matt Pocock's skills for more rigorous agentic engineering now that my code base is climbing towards 100kloc, but can't yet testify to their effectiveness. People I trust swear by them:
https://github.com/mattpocock/skills