Better Than Vibe Coding: Agile AI Driven Development for Complex Apps
Channel: BMad Code (Brian) Format: YouTube video (Part 1 of a series) Published: 2025-04-07 Sponsor: None disclosed
Summary
Brian, a 20-year software veteran, introduces the BMAD Method (Agile-AI Driven Development) as an antidote to vibe coding. Vibe coding works for toys but fails on complex apps: agents go off the rails, credits burn, fixes break previous code. BMAD maps six Agile personas (BA, PM, Architect, PO, Scrum Master, Dev) onto AI roles, doing all the planning outside the coding IDE so the dev agent only consumes one well-scoped story per fresh chat thread. GitHub: bmadcode/BMAD-METHOD.
See bmad-method for the full concept page.
Key Points
- Vibe coding fails on complex apps — fine for play (space shooter, Super Whisper riffing), but for serious projects you need workflows and stability
- BA → PM → Architect → PO → SM → Dev — six sequential AI persona handoffs, each producing concrete markdown artifacts
- Steps 1–5 happen outside the IDE in free/cheap LLM chats with deep research modes — saves credits before Cursor/Windsurf opens
- PRD from PM, architecture doc from Architect (full stack, libraries, schemas, deployment), task list from PO, epics + stories from SM
- Each story is fully self-contained: data models, file locations, project structure, prior-story context — dev agent handles one story per fresh chat thread
- Enabler stories for setup tasks (account creation, env vars, secrets) — some doable by agent, some by human, but laid out explicitly
- 80–90% test coverage target; tests written per story to prevent later regressions
- Brian deliberately avoids “complex prompts and complex rules” — the story is the prompt
Sponsorship & Bias Notes
Sponsor: None disclosed.
Product placement / affiliations: BMAD-METHOD is Brian’s own GitHub project — the video is essentially a soft launch/explainer. Disclosed by Brian himself, not hidden.
Comparison bias: Brian’s claim that BMAD works “in industry, and I’ve tested this” is unverified — single-creator method, no third-party case studies in this video.
Connected Pages
- bmad-method — full concept
- bmad-code — author hub
- ai-coding-workflow — Cole Medin’s parallel/lighter pattern
- four-prompting-disciplines — broader theory of specification engineering
- five-levels-of-ai-coding — BMAD sits at L3–L4
See Also
- Cole Medin’s full AI coding process — adjacent and lighter
- Four Prompting Disciplines — the theoretical frame