A practical guide to the Breakthrough Method of Agile AI-Driven Development. Each lesson covers a phase, persona, artifact, or grooming pattern — with a workflow diagram and a ready-to-use prompt template.
Artifact-driven AI development
BMAD (Breakthrough Method of Agile AI-Driven Development) treats documentation as the primary source of truth. Every AI session starts from a shared artifact chain, not a blank prompt — solving the context amnesia problem that makes unstructured AI development unreliable.
Analysis → Planning → Solutioning → Implementation
BMAD organises every project into four sequential phases, each producing specific artifacts that feed the next. Work cannot begin in a phase until the prior phase's artifacts exist and have been reviewed by a human.
Quick Flow · BMad Method · Enterprise
Not every change needs the same depth of process. BMAD routes work to one of three tracks — Quick Flow, BMad Method, or Enterprise — based on scope, risk, and coordination requirements. The same routing logic also distinguishes greenfield from brownfield projects.
Defining the problem space
The BA persona opens every BMAD project by turning a vague problem into a structured, reviewable artifact. Its sole output — the one-page PRD — must be complete and human-approved before the PM persona begins.
Stories, criteria, and scope
The PM persona takes the approved PRD and produces a prioritised set of user stories, each with explicit acceptance criteria covering both success and failure paths. A story without complete AC is not ready for implementation.
Minimum viable design decisions
The Architect produces the minimum structural decisions needed to start implementation without making choices that are expensive to reverse later. Not a comprehensive system diagram — just the decisions that would diverge across sessions if left implicit.
Implementation against spec
The Developer persona implements stories one at a time, working strictly from the approved artifact chain. Its output is code and tests that satisfy the acceptance criteria — not an interpretation of them.
Backlog, sequencing, and flow
The Scrum Master persona takes the approved story list and sequences it into a deliverable backlog, identifying dependencies, blocking risks, and the minimum viable increment that can be shipped first.
Quality at every stage, not the end
BMAD has no explicit QA persona. Instead, each persona applies a quality gate to its own output before handing off. Quality in AI-driven development is a prevention model, not a post-implementation catch model.
Constraints force clarity
The PRD is the first artifact in every BMAD project. Its one-page constraint is not a formatting choice — it is a forcing function that prevents work from beginning until the problem is genuinely understood.
Testable, not vague
BMAD user stories are sized for a single AI session and paired with acceptance criteria that cover both success and failure paths. A story without complete AC is not ready for implementation.
Decisions that travel with the work
ADRs are the Architect persona's primary output. Each record captures a single design decision, the options considered, the rationale, and the consequences — making reasoning auditable and implementation predictable across sessions.
AI-executable specifications
An implementation plan translates a user story and architecture constraints into a concrete sequence of steps that an AI can execute in a single session without making unreviewed decisions.
Why AI-driven sprints still need refinement
Story grooming in BMAD takes raw stories from the PM phase and makes them genuinely ready for implementation — unambiguous, correctly sized, and free of any interpretation choice the AI would otherwise make silently.
All personas in the room at once
Party Mode runs multiple BMAD personas simultaneously against a story or PRD, surfacing cross-domain conflicts and gaps that single-persona review would miss. It is the closest AI equivalent to a sprint planning meeting.
One persona at a time, focused review
Sequential Grooming passes a story through each relevant persona in a fixed sequence, each performing a focused domain review. Less expensive than Party Mode and suitable for stories where the likely ambiguity is narrow.
Making stories small enough to ship
Scope Hammering is the BMAD grooming technique for stories too large for a single AI session. The PM persona actively shrinks story scope to the minimum deliverable unit that still produces testable value.
Drawing out requirements stakeholders cannot articulate
Advanced Elicitation is the BMAD grooming pattern for stories where stakeholders know what outcome they want but cannot specify what to build. Socratic questioning, negative space probing, and example-driven techniques replace traditional requirements gathering.
Generating options before committing to an approach
BMAD Brainstorming is a structured diverge-then-converge session where multiple personas generate solution options before the team commits to any one approach. It prevents premature lock-in and surfaces alternatives the original author did not consider.
Red-teaming stories and designs before implementation
Adversarial Review is a BMAD grooming pattern where one persona is explicitly activated to challenge, break, and find failure modes in an artifact produced by another. It is the structured equivalent of a red team and prevents the optimism bias that single-persona review produces.
Lightweight grooming for well-understood stories
Quick Dev is the minimal grooming path for stories that are already well-specified, low-risk, and constrained to a single clear implementation. It bypasses the full grooming ritual and moves directly to implementation with a single readiness check.
Technical design as a grooming step
Solutioning is the grooming pattern where the Architect persona produces a lightweight technical design for a story before implementation begins. It prevents the Developer persona from making uncontrolled architectural decisions mid-implementation.
Different constraints, different artifact chains
BMAD distinguishes greenfield projects (new codebase, no inherited constraints) from brownfield projects (existing codebase, technical debt, integration points). The artifact chain and the Architect's responsibilities differ significantly between the two.
Why AI forgets and how artifacts fix it
AI assistants have no persistent memory between sessions. BMAD's artifact chain is designed specifically to solve this: each session begins by loading the relevant artifacts, giving the AI the context it needs to continue coherently.
When to use each, when to combine them
BMAD and SpecOps are complementary methodologies for AI-driven development. BMAD structures the process; SpecOps structures the output. The right choice depends on how well-understood requirements are when work begins.