AI Assisted Coding: Beyond AI Code Assistants: How Moldable Development Answers Questions AI Can't With Tudor Girba In this BONUS episode, we explore Moldable Development with Tudor Girba, CEO of feenk.com and creator of the Glamorous Toolkit. We dive into why developers spend over 50% of their time reading code—not because they want to, but because they lack the answers they need. Tudor shares how building contextual tools can transform software development, making systems truly understandable and enabling decisions at the speed of thought. The Hidden System: A Telco's Three-Year Quest "They had a system consisting of five boxes, but they could only enumerate four. If this is your level of awareness about what is reality around you, you have almost no chance of systematically affecting that reality." Tudor opens with a striking case study from a telecommunications company that spent three years and hundreds of person-years trying to optimize a data pipeline. Despite massive effort and executive mandate, the pipeline still took exactly one day to process data—no improvement whatsoever. When Tudor's team investigated, they asked for an architecture diagram. The team drew four boxes representing their system. But when Tudor's team started building tools to mirror this architecture back from the actual code, they discovered something shocking: there was an entire fifth system between the first and second boxes that nobody knew existed. This missing system was likely the bottleneck they'd been trying to optimize for three years. Why Reading Code Doesn't Scale "Developers spend more than 50% of their time reading code. The problem is that our systems are typically larger than anyone can read, and by the time you finish reading, the system has already changed many times." The real issue isn't the time spent reading—it's that reading is the most manual, least scalable way to extract information from systems. When developers read code, they're actually trying to answer questions so they can make decisions. But a 250,000-line system would take one person-month to read at high speed, and the system changes constantly during that time. This means everything you learned yesterday becomes merely a hypothesis, not a reliable answer. The fundamental problem is that we cannot perceive anything in a software system except through tools, yet we've never made how we read code an explicit, optimizable activity. The Context Problem: Why Generic Tools Fail "Software is highly contextual, which means we can predict classes of problems people will have, but we cannot predict specific problems people will have." Tudor draws a powerful parallel with testing. Nobody downloads unit tests from the web and applies them to their system—that would be absurd. Instead, we download test frameworks and build tests contextually for our specific system, encoding what's valuable about our particular business logic. Yet for almost everything else in software development, we download generic tools and expect them to work. This is why teams have tens of thousands of static analysis warnings they ignore, while a single failing test stops deployment. The test encodes contextual value; the generic warning doesn't. Moldable Development extends this principle: every question about your system should be answered by a contextual tool you build for that specific question. Tools That Mirror Your Mental Model "Whatever you draw on the whiteboard—that's your mental model. But as soon as the system exists, we want the system to mirror you back that thing. We make it the job of the system to show our mental model back to us." When someone draws an architecture diagram on a whiteboard, they're not documenting the system—they're documenting their beliefs about the system. The diagram represents wishes when drawn before the system exists, but beliefs when drawn after. Moldable Development flips this: instead of humans reading code and creating approximations, the system itself generates the visualization directly from the actual code. This eliminates the layers of belief and inference. Whether you're looking at high-level architecture, data lineage across multiple technologies, performance bottlenecks, or business domain structure, you build small tools that extract and present exactly the information you need from the system as it actually is. The Test-Driven Development Parallel "Testing was a way to find some kind of class of answers. But there are many other questions we have, and the question is: is there a systematic way to approach arbitrary questions?" Tudor explains that Moldable Development applies test-driven development principles to all forms of system understanding. Just as we write tests after we understand the functionality we need, we build visualization and analysis tools after we understand the questions we need answered. Both approaches share key characteristics: they're built contextually for the specific system, created by developers during development, and composed of many small tools that collectively model the system. The difference is that TDD focuses on functional decomposition and known expectations, while Moldable Development addresses architecture, security, domain structure, performance, and any other perspective where functional tests aren't the most useful decomposition. From Thousands of Features to Thousands of Tools "In my development environment, I don't have features. I have thousands of tools that coexist. Development environments should be focused not on what exists out of the box, but on how quickly you can create a contextual tool." Traditional development environments offer dozens of features—buttons, plugins, generic views. But Moldable Development environments contain thousands of micro-tools, each answering a specific question about a specific system. The key is making these tools composable and fast to create. Rather than building monolithic tools that try to handle every scenario, you build small inspectors that show one perspective on one object or concept. These inspectors chain together naturally as you drill down from high-level questions to detailed investigations. You might have one inspector showing test failures grouped by exception type, another showing PDF document comparisons, another showing cluster performance, and another showing memory usage—all coexisting and available when needed. The Real Bottleneck To Learning A System: Time to the Next Question "Once you do this, you will see that the interesting bottleneck is in the time to the next interesting question. This is by far the most interesting place to be spending energy." When you commoditize access to answers through contextual tools, something remarkable happens: the bottleneck shifts from getting answers to asking better questions. Right now, because answers come so slowly through manual reading and analysis, we rarely exercise the skill of formulating good questions. We make decisions based on gut feelings and incomplete data because we can't afford to dig deeper. But when answers arrive at the speed of thought, you can explore, follow hunches, test hypotheses, and develop genuine insight. The conversation between person and system becomes fluid, enabling decision-making based on actual evidence rather than belief. Moldable Development in Practice: The Lifeware Case "They are investing in software engineering as their competitive advantage. They have 150,000 tests that would take 10 days to run on a single machine, but they run them in 16 minutes distributed across AWS." Tudor shares a powerful case study of Lifeware, a life insurance software company that was featured in Kent Beck's "Test-Driven Development by Example" in 2002 with 4,000 tests. Today they have 150,000 tests and have fully adopted Moldable Development as their core practice. Their business model is remarkable: they take data from insurance companies, throw away the old systems, and reverse-engineer new systems by TDD-ing the business—replaying history to produce pixel-identical documents. They've deployed Glamorous Toolkit as their sole development environment across 100+ developers. Their approach demonstrates that Moldable Development isn't just a research concept but a practical competitive advantage that scales to large teams and complex systems. Why AI Doesn't Solve This Problem "When you ask AI, you will get exactly the same kind of answers. The answer comes quickly, but you will not know whether this is accurate, whether this represents the whole thing, and you definitely do not have an explanation as to why the answer is the way it is." In the age of AI code assistants, it might seem like language models could solve the problem of understanding systems. But Tudor explains why they can't. When you ask an AI about your architecture, you get an opinion—fast but unverifiable. Just like asking a developer to draw the architecture on a whiteboard, you receive filtered information without knowing if it's complete or accurate. Moldable Development, by contrast, extracts answers deterministically from the actual system. Software systems have almost no ambiguity in meaning—they're mathematical, not linguistic. We don't need probabilistic interpretation of source code; we need precise extraction and presentation. The tools you build give you not just answers but explanations of how those answers were derived from the actual system state. Scaling Through Language, Not Features "You need a new kind of development environment where the goal is to create tools much quicker. You need some sort of language in which to express development environments." The technical challenge of Moldable Development is enabling thousands of tools to coexist productively. This requires a fundamentally different approach to development environments. Instead of adding features—buttons and menu i
