Why Code Quality Compounds in the AI Age
Matt Pocock ran the specs-to-code workflow himself. Each time he re-ran the compiler without examining the code, the output got worse. The Pragmatic Programmer calls this software entropy: every change that ignores the whole-system design makes the codebase harder to modify. Pocock’s conclusion inverts the “code is cheap” premise. Bad code is now the most expensive it has ever been, because a hard-to-change codebase blocks every productivity gain AI can offer. Good codebases, by contrast, let AI perform well.
Building Shared Understanding Before Writing a Line
Pocock’s fix for misaligned output is a two-line prompt he calls “grill me”: “Interview me relentlessly about every aspect of this plan until we reach a shared understanding.” The repo hit 13,000 stars. The prompt generates 40 to 100 questions before the AI considers itself aligned. That conversation then feeds a product requirements document or a list of issues for an async agent. The second technique borrows from domain-driven design: a ubiquitous language file, scanned from the codebase, gives both you and the AI a shared vocabulary that tightens planning and shortens the thinking trace.
Feedback Loops and Test-Driven Development
Even with TypeScript, browser access, and automated tests wired in, the LLM defaults to generating large chunks before running any checks. The Pragmatic Programmer calls this outrunning your headlights: the rate of feedback is your speed limit. Pocock’s fix is test-driven development. Writing a test first forces the LLM to take one small step, make it pass, then refactor. Testing is still genuinely hard (unit size, what to mock, which behaviors matter), but a well-designed codebase makes those decisions obvious and keeps the LLM within its feedback window.
Deep Modules: Structuring Code the AI Can Navigate
John Ousterhout’s deep module packs lots of functionality behind a simple interface. Shallow modules invert that ratio: many small files with complex interfaces, and AI handles them poorly because it reads the wrong file or misses a dependency. Pocock’s “improve codebase architecture” skill groups related code behind a clean boundary, then lets AI fill the implementation. You design the interface; AI writes what’s inside. Reviewing only the interface keeps cognitive load low enough that developers stop feeling exhausted, and the result is a codebase that rewards TDD.
Notable Quotes
software fundamentals matter now more Matt Pocock · ▶ 0:34
a concept of a ubiquitous language. Matt Pocock · ▶ 8:26
And so code is not cheap. That’s the Matt Pocock · ▶ 17:16
Key Takeaways
- Specs-to-code workflows degrade codebases by skipping design investment on every iteration.
- The ‘grill me’ prompt forces 40-100 clarifying questions to build a shared design concept before coding starts.
- Deep modules with simple interfaces make codebases testable and easier for AI to reason about correctly.