How To - I Broke BMAD (Again) — But This Time the Agents Saved the Project

When a “quick cleanup” turns your BMAD install into a mess, what do you do next? In this episode, I walk through how I broke BMAD in a brownfield repo, fixed it the hard way, and then pushed the experiment forward using Windsurf, Claude Sonnet 4.6, and BMAD’s personas on a multimodal neural network project. ​ From Copilot chaos to Windsurf sanity Recapping how Copilot-driven “syntax cleanups” and piecemeal PRs tangled BMAD config files and exhausted my personal Copilot credits. ​ Moving the work to Windsurf on my work laptop (via RustDesk) to use company-provided AI capacity instead of burning my own allocation. ​ Cleaning up the BMAD mess Showing the trail of commits, sub-PRs, and reinstall attempts that actually made the BMAD setup worse before it got better. ​ Explaining why one-by-one “fix every warning” commits are risky, and how a bulk, focused PR would have been safer. ​ Getting the codebase and BMAD install back into a healthy state so the brownfield experiment could continue. ​ Letting BMAD sharpen the PRD Using BMAD’s Analyst persona to reassess and upgrade the original PRD (initially drafted with Perplexity) for the multimodal neural network. ​ Strengthening the problem statement, target user (independent AI devs and hobbyists), and key differentiators like “no cloud required” and competitive landscape. ​ Highlighting how BMAD can periodically re-validate assumptions in a long-lived codebase, not just generate docs for a greenfield project. ​ Revisiting Agile in an AI world Arguing that generative AI plus BMAD changes the “math” behind the Agile Manifesto tradeoffs between documentation and working software. ​ Suggesting that AI drastically lowers the cost of things on the “right side” (docs, plans, specs) without sacrificing individuals, collaboration, or responding to change. ​ Making the case that the bottleneck has shifted from “how fast devs can build” to “how fast the business side can feed good context into agents.” ​ Party mode with agents and new stories Spinning up BMAD’s party mode with Mary (Analyst), Amelia (Dev), Winston (Architect), Murat (Test Architect), Quinn (Tester), John (PM), and Bob (Scrum Master). ​ Asking the agents to validate epics and stories, call out gaps, and then auto-generate 26 standalone user story files from an epic so development can proceed. ​ Positioning the project as research-focused: BF16 training, gradient checkpointing, flash attention, ablation studies, and math reasoning benchmarks aimed at publishable results. ​ Lessons from screwing up in public Owning the frustration, confusion, and embarrassment of breaking BMAD, and using it as a real-world example of AI-assisted dev pain. ​ Emphasizing that agents aren’t at fault when constraints and intent are underspecified—the responsibility still sits with humans. ​ Encouraging viewers to experiment, build AI fluency, and treat these tools like over-eager junior teammates that still need clear direction. ​ If you’ve ever wondered what happens when BMAD, AI assistants, and human error collide in a real brownfield repo—and how to recover while still moving a serious research project forward—this episode gives you the full, unsanitized story, plus a glimpse of where agentic development can take us next.