How To - BMAD vs. My Old Code: Dropping an AI Dev Framework into a Real Brownfield Repo

Bringing AI agents into a greenfield repo is one thing; dropping BMAD into a messy, already-built “real” project is something else entirely. This episode walks through how to add the BMAD method v6 to a brownfield multimodal neural network repo and what happens next. ​ Why this brownfield matters Using one of Tim’s most complex personal projects: a multimodal neural network designed to run on consumer and edge hardware (CPUs, GPUs, NPUs, TPUs) instead of giant data-center rigs. ​ Originating from a client conversation about models on low-power edge devices and a reaction to the “go big or go home” LLM arms race described in Karen Ho’s “Empire of AI.” ​ Framing the repo as an experiment in building small, capable models that regular people can run without hyperscaler infrastructure. ​ The original brownfield build Started in November 2025 using Perplexity to help draft a product requirements document, then GitHub Copilot to generate and evolve the code over a couple of intense weekends. ​ Targeting academic researchers, independent AI developers, hobbyists, and small orgs as an open-source proof-of-concept rather than a commercial product. ​ Carefully documenting hardware assumptions (RTX 4080, Apple Silicon, AMD/Intel machines), training pipelines, project structure, and best practices (testing, CI/CD, security). ​ Installing BMAD v6 into the brownfield repo Demonstrating a full BMAD v6 install via npm into the new brownfield clone, including core, Agile, Builder, and Creative Intelligence modules. ​ Wiring in multiple agent backends (Copilot, RooCode, Cursor, Windsurf, Claude/Codacy, etc.) while watching Copilot tokens climb to 99.4% for the month. ​ Confirming all BMAD directories, agents, and integration configs land correctly alongside the existing project structure. ​ Letting BMAD analyze the existing system Asking Mary, the BMAD business analyst persona (via RooCode), to review the original PRD and generate a structured assessment. ​ Getting high marks on technical quality and risk management, with clear critique around business clarity, market viability, and go-to-market thinking. ​ Surfacing gaps: competitive analysis, user prioritization, ROI justification, sustainability and governance models, and user journey mapping. ​ Having Mary generate a written PRD assessment summary and committing it back into the repo as a starting artifact for future BMAD workflows. ​ What this experiment is really about Using BMAD v6 to “deconstruct” a hand-built, Copilot-driven brownfield repo—much like a real product team would do—rather than only showing clean greenfield demos. ​ Treating this as a public lab: inviting others to explore the repo, see the BMAD workflows and personas in action, and adapt the approach to their own brownfield projects. ​ Previewing next steps: running full BMAD workflows (analyst, PM, architect, dev, party mode) over the brownfield code to see how far the method can go in refactoring and improving it. ​ If you’ve ever wondered how to bring an opinionated agentic framework like BMAD into an already-existing codebase—without rewriting everything from scratch—this episode is your long-form, over-the-shoulder walkthrough, brought to you by Tim Unscripted.