Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
⚡️ Matt Pocock - Why Engineering Fundamentals matter MORE now
TL;DR
Engineering fundamentals matter more in the AI era, not less — Matt Pocock says every time he treated code as a disposable “compile target” for English prompts, he got a mess, while classic ideas from The Pragmatic Programmer, Extreme Programming, DDD, and A Philosophy of Software Design made AI-generated code far easier to steer.
The best AI coding workflows depend on clean interfaces and deep modules — Pocock argues that if a codebase is easy for humans to change, it’s easier for AI too, and he points to John Ousterhout’s “deep modules” idea: lots of functionality behind a simple, intentional interface.
Domain-driven design is suddenly useful again because models already understand it — instead of inventing fresh terminology, Pocock says you can use DDD’s existing language and structures so that “the AI and you are speaking the same language,” especially through tools like ubiquitous-language docs.
His most practical AI trick is a shared vocabulary file inside the repo — one of his mattpocock/skills prompts, which has roughly 13,000 GitHub stars, scans a codebase to build a markdown glossary of domain terms that he keeps open while prompting for clarity and consistency.
People still mostly want traditional teaching, even in AI — despite experimentation with interactive formats, Pocock says lectures still work best for most learners, with his latest two-week cohort distilled from 2 months of prep, 4.5 hours of edited video, and about 100 lesson units.
AI tooling is converging, and the next big trade-off is control versus convenience — Pocock thinks coding agents are settling on similar patterns, and he’s now interested in “agent-agnostic” software-factory methods and the inversion-of-control question: boxed systems like Claude Code are easy, while primitive-based systems like Pi offer observability and control.
The Breakdown
From Total TypeScript to AI Hero
The conversation opens with Matt Pocock in London, joking about the event being “the worst I’ve ever been to” before quickly getting into how his work has shifted from Total TypeScript to running AI Hero. He now teaches AI in two-week cohorts because the space moves too fast for evergreen self-paced courses — and, in classic Matt fashion, major product changes keep landing right as his courses begin.
Why old-school software books suddenly feel new again
Pocock’s big thesis is that the hot take — “code doesn’t matter anymore” — just hasn’t held up in practice. Whenever he tried to ignore the code and treat AI like a pure English-to-software compiler, he ended up with a “terrible mess,” so he went back to older texts like Extreme Programming, The Pragmatic Programmer, DDD, and A Philosophy of Software Design and started pulling those ideas directly into prompts.
Narrow waists, deep modules, and containing the slop
The hosts riff on architecture patterns that matter even more with AI: narrow waists, MVC, ports-and-adapters, and simple interfaces around messy internals. Pocock connects this to John Ousterhout’s “deep modules” idea — lots of capability behind a clean surface — and says that’s exactly where AI shines: humans define the interface intentionally, then delegate the implementation inside the box.
DDD as a language both humans and models already share
On domain-driven design, Pocock says the appeal is that it’s already a mature, composable framework living in model “latent space.” Instead of inventing brand-new abstractions, you can use DDD’s building blocks so the model understands what you mean, and he frames the goal neatly: align the code with language so that you and the AI are operating from the same domain model.
The markdown glossary trick he keeps open while prompting
He makes that concrete with one of his most popular repo skills — mattpocock/skills, which he says has around 13,000 stars. The “ubiquitous language” skill scans a codebase, finds domain-specific jargon, and produces a markdown file he keeps open while prompting, so terms like “mole” resolve to the right meaning — health app, spy, or zoo animal — instead of drifting.
Why he’s unusually good at explaining technical things
Asked about his teaching style, Pocock says he spent 6 years as a voice coach before becoming a developer, and that communication has always felt like an unfair advantage. His process is obsessive and structured: for his latest course he spent about 2 months collecting notes in Obsidian, grouping ideas, prioritizing lessons as P1/P2/P3, then building around the essentials until only the strongest material survived.
What learners actually want: mostly lectures
He breaks teaching into knowledge, skills, and wisdom, with wisdom being the hardest to teach and the place where AI can help through dialogue. But the surprising thing, he says, is that the more experimental and AI-native he gets with format, the more people bounce off it — most learners still want the old-school setup of lectures, maybe some exercises, and not much more.
TypeScript, agent convergence, and life beyond coding
Near the end, the talk widens out: TypeScript seems increasingly central to AI app building because of its ecosystem and frontend-heavy product work, even if Pocock avoids making grand predictions. He’s now focused on agent-agnostic methods, the trade-off between convenience and observability in systems like Claude Code versus lower-level primitives like Pi, and portable “skills” like his tiny “grill me” prompt that forces an AI to question assumptions until it reaches a shared understanding with you.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
