AI Coding Workflows with Paul Mason

Tim Williams: Hey everyone, welcome to the very first episode of this new developer podcast. Tim Williams: Uh, I'm Tim Williams, Lead Developer of a small but mighty team that builds everything from web apps to internal tools to, you know, the occasional experiment that probably should've stayed in staging. Tim Williams: Over the years I've worked across more stacks than I can count — AWS, Node, React, Vue, PHP — and what's kept me hooked isn't just the code, it's the craft behind it. Tim Williams: This show's my way of exploring that craft in the modern era where AI now sits right beside us in the editor. Tim Williams: We'll talk about tools, techniques, leadership, and the human side of software development. Tim Williams: The goal's to keep things honest, useful, and, uh, maybe a little funny because developers need both precision and perspective to stay sane. Tim Williams: Whether you're debugging at two in the morning or mentoring someone through their first merge conflict, this podcast should feel like sitting down with people who get it. Tim Williams: For our first episode I'm joined by Paul Mason, a full-stack developer from Seattle, Washington. Tim Williams: Paul's been deep into AI-assisted workflows for a while now, especially with Cursor. Tim Williams: We're gonna talk about how to get the most out of it — when to switch between Ask, Plan, and Agent — and why "trust but verify" might be the most important rule in modern coding. Tim Williams: Paul, man, welcome to the show. Paul Mason: Thanks Tim, yeah, super excited to be here. Paul Mason: I've been following your writing for a while and, uh, this feels like the right time for this kind of conversation. Paul Mason: Every dev I know is still figuring out how AI fits into their daily workflow. Tim Williams: Yeah, totally. Before we dive into the workflow stuff, let's, uh, do a quick state of AI coding tools. Tim Williams: On October 7th 2025, Cursor launched its new Plan Mode, which lets the model research your codebase, build an editable plan, and run longer agents for complex changes. Tim Williams: That move basically made planning a first-class citizen instead of an afterthought. Paul Mason: Then on May 19th 2025, GitHub announced the public preview of its Copilot Coding Agent. Paul Mason: That one goes beyond autocomplete — it can actually implement tasks end-to-end and even handle pull requests. Paul Mason: Pretty wild seeing that land in production IDEs. Tim Williams: Right. And back in January 2025, JetBrains launched "Junie," its own AI coding agent, bringing full agentic workflows into IntelliJ and PyCharm. Tim Williams: So yeah, the arms race is on. Tim Williams: Sourcegraph's pushing Amp as a production-grade coding agent, Windsurf's evolving the whole "agentic IDE" thing, and Amazon Q Developer keeps layering in reviews and doc generation right inside the editor. Paul Mason: Replit's got its own Agent now too — it plans, builds, tests, and even self-fixes stuff in the browser. Paul Mason: So, uh, basically every major tool's converging on the same idea: plan first, execute, then verify while keeping a human in the loop. Tim Williams: Exactly. So let's ground a bit on Cursor itself. Tim Williams: Plan Mode lets you preview the steps before anything happens, which is, I think, huge for bigger edits. Tim Williams: Browser controls and longer-running agents help with UI debugging and those annoying end-to-end tasks that used to eat afternoons. Paul Mason: Yeah, I like that Cursor sort of nudges you to think before acting. Paul Mason: You get a plan, you edit it, make it yours, and then you hit go. Tim Williams: Cursor feels like magic at first, right? Tim Williams: You type something vague and, boom, there's working code. Tim Williams: But the magic fades fast if you don't actually know what changed or why. Paul Mason: Totally. Paul Mason: Think of Ask as talking the problem through, Plan as deciding what to do, and Agent as the "okay, let's ship this" phase. Paul Mason: You need all three if you want speed without losing control. Tim Williams: Yeah, that's such a good way to put it. Tim Williams: Most people live in Agent mode because it feels productive — you see code appear and you get that dopamine hit. Tim Williams: But it's like driving in fifth gear through a parking lot — looks fast, ends messy. Paul Mason: Yeah, I've done that. Paul Mason: So a few weeks ago I built this reporting function that merged data from two APIs. Paul Mason: I started in Ask mode just to get my head straight, like, "Okay, how do I normalize inconsistent fields without blowing up performance?" Paul Mason: Then I switched to Plan. Paul Mason: The model laid out a nice flow — fetch, sanitize, normalize, merge — but, uh, I noticed it skipped error handling and some production quirks. Paul Mason: So I tweaked the plan, added validation, asked Cursor to review it, and it actually flagged a naming collision I'd missed. Paul Mason: Once the plan looked solid, I flipped to Agent. Paul Mason: It wrote maybe eighty percent of the function, and I filled in the rest — cleaned logic, added logs, ran tests. Paul Mason: First run passed. Paul Mason: I think the whole loop — Ask, Plan, Refine, Execute — cut my time in half. Tim Williams: That's awesome. Tim Williams: You stayed in control the whole time, using the tool like a fast assistant instead of an autopilot. Tim Williams: That's, uh, exactly how it's meant to work. Tim Williams: Yeah, yeah. Paul Mason: But we should be honest — there's a lot of junk out there right now. Paul Mason: AI slop, as people call it — half-baked tutorials, broken code, content that just feels synthetic. Paul Mason: I get why some devs are, you know, AI-phobic. Tim Williams: Yeah, I think they have a point. Tim Williams: There's an ocean of low-effort material flooding the web, and it's exhausting. Tim Williams: But it's not the tech that's the problem — it's the way people are using it. Tim Williams: Every wave of automation starts messy before it finds its discipline. Tim Williams: The way I picture it is kind of like the Renaissance workshops. Tim Williams: Michelangelo and others trained apprentices to paint the easy parts — the skies, the fabrics, the backgrounds. Tim Williams: Then the master came back to paint the faces, the hands — the parts that carried emotion. Tim Williams: The workshop produced the painting, but the master owned the quality. Paul Mason: Yeah, that's exactly how I see AI in code. Paul Mason: Let it paint the background, but you still handle the faces. Paul Mason: The art lives in the details you touch yourself. Tim Williams: Totally. Tim Williams: AI can handle the boilerplate, but it can't feel context. Tim Williams: The developer's still the one shaping the meaning, the architecture, the little human touches that make it elegant. Tim Williams: "Trust but verify" sounds old-school, but man, it fits perfectly here. Tim Williams: I review every agent diff like it came from a new hire and sometimes even ask the model to explain its own changes. Tim Williams: If it can't tell me why, that's my cue to dig in. Paul Mason: Same here. Paul Mason: If I can't explain a change, it doesn't merge. Paul Mason: That rule's saved me from, uh, way too many weekend rollbacks. Tim Williams: Been there. Tim Williams: And it's wild how many subtle bugs you catch just by asking, "Why did you do that?" Tim Williams: And just so this doesn't sound like a Cursor commercial — they're not alone. Tim Williams: GitHub Copilot's Agent came out May 19 2025, JetBrains' Junie landed January 23 2025, and now you've got Sourcegraph Amp, Windsurf, Amazon Q, and Replit's Agent all marching in the same direction. Tim Williams: Everyone's learning the same lesson: plan first, supervise execution, verify results, and document your reasoning. Paul Mason: Yeah, it's convergence. Paul Mason: And honestly, that's healthy. Paul Mason: It means we're, uh, standardizing on workflows that keep humans in charge. Tim Williams: So, three quick takeaways. Tim Williams: Use Ask to clarify your thinking before you code. Tim Williams: Use Plan to make your intentions visible and editable. Tim Williams: And when you switch to Agent, stay alert — review every diff like your name's on the line. Paul Mason: I'd add one more. Paul Mason: Leave breadcrumbs — comments, notes, why you accepted or changed a suggestion. Paul Mason: Future you will, uh, absolutely thank present you. Tim Williams: Yeah, self-documentation is an act of kindness. Tim Williams: Paul, thanks for kicking off episode one with me. Tim Williams: This sets the tone — thoughtful, practical, and human. Paul Mason: Thanks Tim, this was a blast. Paul Mason: It's nice to have a space that takes the hype out of AI and puts the craft back in. Tim Williams: Couldn't agree more. Tim Williams: And to everyone listening, thanks for hanging out with us for episode one. Tim Williams: We'll be back soon with more conversations about the evolving world of software — the tools, the people, and that balance between trust and verification. Tim Williams: Until then, keep building, keep questioning, and, uh, keep your commits clean.