Geoff Mazeroff

AI

Top AI Trends From 100 Interviews | AI Briefings: nWave

• nWave runs on Claude. It helps for the whole SDLC, and is based on principles and fundamentals. It has (at time of writing) 23 agents and 90+ skills.
• Engineers need to understand the problem space for thinking about the solution space.
• Our job is to describe in verifiable detail the nature of the problem we’re solving.
• For simple one-off applications, vibe coding works absolutely fine. But if your dealing with systems that handle other people’s money or private information, the broader of context of consideration matters much more.
• Most of the code AI has seen is likely small, simple, and relatively naive because that’s what most software in public looks like.
• Working iteratively is just as important with AI tools. Iteration is part of the process for discovery of the problem.

Software engineering

It Ain’t Broke: Why Software Fundamentals Matter More Than Ever

• There’s a “specs to code” movement where you write the specs, AI writes the code. If the code isn’t correct, just change the specs; however this usually results in more problems. “It’s just vibe coding by another name.”
• Bad codebase = hard to change.
• Software entropy — code quality gets worse overtime (aka: software rot).
• False premise: Code is cheap. Bad code is more expensive with AI.
• Good codebases make it easier to get better value out of AI.
• Failure mode 1: The AI didn’t do what I wanted. But no one knows exactly what they want (imperfect language).
• Tip 1: /grill-me skill: Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one-by-one.
• Failure mode 2: AI is way too verbose.
• Tip 2: Use Eric Evans’ concept of ubiquitous language. Scan the codebase for terminology then create a markdown file.
• Failure mode 3: AI produced code that doesn’t work.
• Tip 3: Put in place feedback loops so AI can know it didn’t work (static types, browser access for web apps, automated tests).
• Failure mode 4: AI does way too much.
• Tip 4: Take small, deliberate steps. The rate of feedback is your speed limit. Use TDD.
• Failure mode 5: AI doesn’t understand my code.
• Tip 5: To make your code easier to test, your system should have “deep modules” with lots of functionality behind a simple interface. Improve your codebase architecture to deepen modules.
• Failure mode 6: My brain can’t keep up.
• Tip 6: Design the interface, delegate the implementation. You need to be aware of modules.
  Code is not cheap after all. Your job is to think strategically if the AI is thinking tactically.
• https://github.com/mattpocock/skills