The Software Factory Still Needs a Human

The Software Factory Still Needs a Human

Everyone wants a software factory: tickets go in, reviewed and tested code comes out. So how do you actually build one, and where should the human stay? Brendan and Baird walk through the setup they already use on a side project. A recorded planning call becomes a spec, the spec becomes Linear tickets, and Claude Code picks the tickets up. Then they look at what's still missing: automatic pickup, real checks and balances, and a way to run it all without a $145 API bill for a routine task. They agree the code itself can go to the AI. The human still decides what gets built and signs off on how it works. Baird argues for small chunks over one-shot builds, because AI over-engineers and you lose track of what you have. He also makes the case for feeding Granola meeting transcripts straight back into the tickets. The back half covers why you should dogfood what you build ("seven apps, two users"), the idea that you can give up context but not understanding, and how vibe-coded prototypes run into trouble when they go to production because the AI follows your bad patterns without pushing back. Also: context windows that quietly make agents dumber, where the SDLC fits, and the two questions to ask your agent before any PR: "Did we do everything we set out to do? Did we do anything extra?" Want help building your own factory, or moving a vibe-coded app to production? We're taking consulting work. Drop a comment or DM us. Sponsor Powerful podcast hosting & analytics By podcasters, for podcasters. Since 2016 Brought to you by https://fireside.fm claudecode #claudeai #aicoding #aiagents Signup for our newsletter: https://vibing-with-ai-code.beehiiv.com Check out our Website https://vibingwithaicode.com

Your AI Assistant Should Act, Not Just Report

Six months into running his Claude Command Center, Brendan has a confession: the morning brief broke for a week and he didn't notice, because he'd already stopped reading it. This episode is the honest post-mortem on why a personal AI system can genuinely make you more productive and still feel like it's missing the point. Sparked by GrokBot and a conversation with Jesse Hanley of Bento, Brendan and Baird dig into what's actually broken — a system that generates content for you to look at but never acts on your behalf, briefs that insist a task isn't done when the proof is sitting in Slack, and the context gaps that make consolidating four channels into one feel worse than juggling four. Baird pushes back on the framing with the better question — are you being productive, or just busy? — and makes the case for scoping agents into real lanes instead of chasing one that does everything: let it audit what you actually do, then build the small utilities and read-only tools it needs to take the first step for you. The back half gets into the guardrails: why production stays read-only no matter how good the models get, why an agent that's right 99 times out of 100 is still a problem, why the "I've been asked this four times" moment is the real automation signal, and the danger nobody talks about — shipping so much cognitive load to the agent that you can't answer why anything was built the way it was. Plus Wade Foster's evening brief, agents-as-employees, and why neither host is ready to let a bot answer a human. Sponsor Powerful podcast hosting & analytics By podcasters, for podcasters. Since 2016 Brought to you by https://fireside.fm claudecode #claudeai #aicoding #aiagents Signup for our newsletter: https://vibing-with-ai-code.beehiiv.com Check out our Website https://vibingwithaicode.com

Building a Personal AI Agent That Cuts Through the Noise

Baird's been building a reactive, background agentic framework — something that sits around all day surfacing what actually matters instead of waiting to be kicked off — and Brendan compares it head-to-head with his Claude Command Center. The whole episode circles the hardest problem in personal AI systems: cutting through the noise. Both hosts are fighting the same battle from different angles — a morning brief that dumps 30 to-dos when only two matter, figuring out whether a Slack message was actually answered, piecing together a Gmail thread and a Slack message that are secretly about the same thing, and Brendan quietly abandoning his own morning briefing because it stopped being useful. The big takeaway that keeps surfacing: the answer isn't "let AI do everything." A lot of this work is mechanical — data processing, cleaning, webhooks and cron — and doing that part in code first, then bringing AI in only where judgment is actually needed, is what makes the whole thing reliable. Plus: Vercel's new Eve framework, why an agency's fragmented context is harder than a single company's, an agent that got so confident it stopped opening PRs and committed straight to main, and how AGENTS.md / CLAUDE.md let a project teach itself as it goes. Sponsor Powerful podcast hosting & analytics By podcasters, for podcasters. Since 2016 Brought to you by https://fireside.fm claudecode #claudeai #aicoding #aiagents Signup for our newsletter: https://vibing-with-ai-code.beehiiv.com Check out our Website https://vibingwithaicode.com

The Plan Matters More Than the Model

With Fable's Pro-plan window closing, Brendan and Baird both did the same thing on their own: instead of trying to build with it, they had Fable plan — full roadmaps and individual feature specs — so cheaper models like Opus and Sonnet could execute against a high-quality plan. That kicks off a wider check-in on where AI is actually landing right now. Models feeling far more thorough and meticulous, Opus writing Python scripts to verify Ruby physics it can't run, Claude Code's improved auto-permissions mode finally letting you kick off work and walk away, and why agent loops still aren't the answer when there's no clear "right" to test against — especially for subjective frontend and design work. They close on the real distinction that's going to matter: AI is a tool that fills gaps you couldn't otherwise fill, but take the human out and all you've got is slop. Sponsor Powerful podcast hosting & analytics By podcasters, for podcasters. Since 2016 Brought to you by https://fireside.fm claudecode #claudeai #aicoding #aiagents Signup for our newsletter: https://vibing-with-ai-code.beehiiv.com Check out our Website https://vibingwithaicode.com

Activation Energy

Brendan dove back into three stalled projects in one day — a personal command center, a video game tracker, and a Rails 8 app he hadn't touched in two months — and both hosts dig into why AI makes restarting old projects so much easier: it remembers where you left off, tells you what's actually built, and (with Fable) plans and cross-checks the reassessment before you write a line of code. They get into when it's worth splitting context (personal vs. project-specific memory), why Fable feels like the best planner but not necessarily the best daily driver, and how tools like Conductor now auto-select models per task. Plus: what actually happens when sub-agents run out of session usage mid-task, and why lowering activation energy might be the single biggest thing AI has changed about how they work. Sponsor Powerful podcast hosting & analytics By podcasters, for podcasters. Since 2016 Brought to you by https://fireside.fm claudecode #claudeai #aicoding #aiagents Signup for our newsletter: https://vibing-with-ai-code.beehiiv.com Check out our Website https://vibingwithaicode.com