BMAD Method (Agile AI-Driven Development)
A structured workflow from BMad Code for building complex applications with AI coding agents. Replaces “vibe coding” with six sequenced AI personas drawn from traditional Agile, each handing artifacts to the next. The point is that all the expensive thinking happens outside the IDE — by the time Cursor or Windsurf opens, the agent only consumes one well-scoped story at a time.
GitHub: bmadcode/BMAD-METHOD
The Six Personas
| Step | Persona | Tool / mode | Output |
|---|---|---|---|
| 1 | Business Analyst | Advanced thinking LLM (chat) | Refined idea, elicited details |
| 2 | Project Manager | Deep research mode (OpenAI / Gemini) | PRD — what to build, MVP scope |
| 3 | Architect | Advanced thinking LLM | Architecture doc — stack, libraries, schemas, deployment |
| 4 | Product Owner | Advanced thinking LLM | Granular, sequenced task list |
| 5 | Scrum Master | Advanced thinking LLM | Epics + stories, each fully self-contained |
| 6 | Developer | Coding IDE (Cursor / Windsurf) | One story per fresh chat thread, with tests |
All artifacts get converted to markdown and dropped into an AAI/ folder for the dev agent to reference.
Why It Works
- Cost discipline — steps 1–5 happen in free or cheap LLM chat windows, not in the IDE where every token burns credits
- Context discipline — each dev story is small enough to fit in one fresh chat thread without overwhelming the LLM
- Test discipline — every story includes tests; target 80–90% coverage as the project grows
- No prompt engineering rabbit hole — Brian explicitly avoids “complex prompts and complex rules for development agents.” The story itself is the prompt.
Story Granularity
Each story is written with everything the dev agent needs: data models from the architecture doc, file locations, project structure, context from the PRD specific to that story, and what’s been done in previous stories. Some stories are enabler stories — e.g., “set up Vercel account, configure env vars, secure secrets” — which may need human action mixed with agent action.
Compared to Other Workflows
- Cole Medin’s PLANNING.md + TASK.md pattern — same goals (plan before coding, golden rules, reference docs in context), but compressed into two markdown files instead of six personas. BMAD is heavier; Cole’s is lighter. Use BMAD for SaaS-scale builds, Cole’s for projects under ~10 stories.
- four-prompting-disciplines — Nate B Jones’s framing of “specification engineering” as the highest discipline. BMAD is one concrete instantiation: the PO + SM artifacts are specifications.
- five-levels-of-ai-coding — BMAD sits around L3–L4: heavy human orchestration of multiple AI roles to ship complex apps.
See Also
- bmad-code — creator
- ai-coding-workflow — Cole Medin’s lighter equivalent
- four-prompting-disciplines — the underlying specification-engineering thesis
- cursor — the dev-agent IDE Brian uses
- Source: Better Than Vibe Coding