How To - Turning a Messy Brownfield Repo into Gold with BMAD, GitHub Copilot, and Claude
Source: Our analysis of the creator's lived experience, based on what they said in this video.
Creator's Key Takeaways
I love that. We never answered. Why does this need to exist? Because it can.
The value prop is the first open-source multimodal you can actually train at home.
I'm being asked to delete items. Why am I being asked to delete these items, though?
This is another how-to video, a BMAD experiment brought to you by Tim.
Creator's Tips & Advice
Questions This Creator Answers
YouTube Video Description↓
Taking BMAD into a brownfield repo is one thing; tightening the product story so it actually makes sense is another. In this how-to episode, we use GitHub Copilot + Claude Sonnet 4.6 and the BMAD Analyst persona to reshape the foundational docs for a multimodal AI project. Picking up after the token crunch Resuming work on the BMAD brownfield repo after Copilot premium tokens reset for the new month. Switching from earlier runs in RooCode with Claude Sonnet 4.5 to GitHub Copilot powered by Claude Sonnet 4.6. Having Mary re‑review the PRD Asking Mary (BMAD’s strategic analyst persona) to reassess the existing Product Requirements Document inside the brownfield repo. Getting a clear verdict: technically implementation-ready, but not ready for stakeholder approval or community review without a stronger value story. Identifying three critical gaps: no clear problem statement, no competitive landscape, and no real go-to-market / launch plan. Co-creating the problem statement and value prop Collaboratively drafting a problem statement around researchers and indie devs paying 200–500 dollars a month for cloud compute and being constrained to unimodal architectures. Clarifying the value proposition: a ~250M parameter multimodal model you can actually train at home on a single consumer GPU, with meta-learning and Wolfram Alpha–backed symbolic reasoning. Positioning “no cloud required” as a core differentiator for independent developers and tinkerers. Adding a real competitive matrix Having Mary build a seven-column competitive landscape matrix comparing this project to other small and open models. Writing the matrix directly into the PRD to support the “no cloud required” narrative with concrete, side-by-side data. Updating the PRD sections (problem statement, target users, differentiators, launch and growth plan) to close the first three gaps from Mary’s earlier assessment. Generating the product brief Answering Mary’s targeted questions: why this repo exists, what a 12-month “win” looks like (e.g., a paper and active GitHub usage), how versions 1.0, 1.5, and 2.0 should scope progression toward edge/IoT production readiness, and whether Wolfram Alpha is core or optional. Choosing a working name—NeuralMix—for the project to make communication and storytelling easier. Having Mary generate a structured product brief from these answers so the next BMAD persona (Architect) can start from stronger business and technical context. A few real-world bumps Encountering an unexpected prompt to delete files, then rerunning the BMAD install via npm to restore missing pieces. Opening and reviewing the pull request that folds the refined PRD and competitive analysis back into the repo. If you’ve ever wondered how to use AI agents not just to write code but to strengthen the story around a complex brownfield project, this episode shows how BMAD’s analyst persona plus Copilot and Claude can help you tighten your PRD and spin up a solid product brief—another BMAD experiment, brought to you by Tim Unscripted