• Matt initially thought AI required throwing away older software practices and starting fresh. 50:54
  • He experimented with spec-driven development and found that repeated spec-to-code loops often made the code worse than hand-coding. 51:21
  • He concluded that agents need good feedback loops just as humans do. 52:13
  • Reading The Pragmatic Programmer pushed him toward concepts like software entropy, tracer bullets, programming by coincidence, and working within feedback loops. 52:33
  • He argues agents can produce software entropy faster than humans because they generate code quickly without long-term strategic judgment. 53:16
  • A tracer bullet means building a thin, working path across layers so feedback arrives early. 53:40
  • Matt saw agents build horizontal layers in isolation, then fail late when the database, application layer, and React components finally met. 53:40
  • He prefers vertical slices because they cross integration layers and reveal whether the system actually makes sense. 54:29
  • A "leading word" is a repeated phrase in a prompt or skill that activates a useful concept in the model and changes its behavior. 54:58
  • The Philosophy of Software Design gave him more agent-useful language, especially "deep modules." 55:30
  • Domain-driven design led him toward building a shared domain language between the human, the codebase, and the agent. 56:53
  • His "Grill with Docs" idea builds domain language during planning so later prompts can be shorter and clearer. 57:17
  • In one app, the term "materialization cascade" captured a complex rule about turning ghost lessons, sections, and courses into real ones. 57:45
  • A codebase with explicit domain terms is easier for an agent to search and navigate. 58:10
  • Matt argues that agents make healthy codebases more important because each session starts fresh, like a new starter with no memory. 1:00:10