bmad-architecture

BMAD Architecture (Solutioning)

Persona: Winston, the Architect. Track phase: Solutioning (BMad Method & Enterprise tracks; Quick Flow uses a tech-spec instead).

Function: Turn the PRD into ONE coherent architecture.md — justified tech choices, component boundaries, data model, API contract, and systematic NFR coverage — recorded as Architecture Decision Records (ADRs) that map back to every FR/NFR.

Why this skill is load-bearing

This is the semantic conflict-prevention layer. Later, the orchestrator fans many parallel dev agents across stories. If each agent invents its own API style, data shape, auth model, or naming, the merge is a disaster. One architecture removes that entire class of conflict in advance:

• API style — REST vs GraphQL vs gRPC, decided once.
• Data model — entities, relationships, ownership, decided once.
• State management — server/client state strategy, decided once.
• Naming & conventions — casing, resource naming, error shape, decided once.
• Security approach — authn/authz model, secrets, decided once.

Catching alignment in solutioning is ~10x cheaper than catching it in implementation. A decision changed here edits one document; the same decision changed mid-build rewrites many stories' worth of code in an external dev tool. Spend the judgment now.

Scope (PLAN, never build)

This skill produces a document. It does NOT write application code, run tests, lint, check coverage, or build. The last artifact is architecture.md (a planning artifact handed to scrum-master / external dev tools). Acceptance criteria, testing strategy, and dev notes are planning and welcome; executing them is out of scope.

Inputs

1. prd.md (required for BMad/Enterprise tracks) — source of FRs and NFRs.
2. project-context.md — the project "constitution" (constraints, existing stack, team size). Load it; respect it.
3. decision-log.md — prior cross-workflow decisions. Read before deciding; append new ADR summaries after.
4. Optional ux-design.md for interface architecture alignment.

Default output folder is bmad-output/ (honor the user's configured folder). Write to bmad-output/architecture.md.

Three intents

Always ask which intent applies if ambiguous; never blindly one-shot.

Create

1. Read prd.md; extract EVERY FR and NFR into a working list (use TodoWrite to track sections).
2. Run the NFR checklist to surface categories the PRD may have under-specified:

   bash ${CLAUDE_PLUGIN_ROOT}/skills/bmad-architecture/scripts/nfr-checklist.sh
   

3. Identify architectural drivers — the NFRs that most constrain design.
4. Pick the architecture pattern matched to the track/scale (don't over-engineer — see REFERENCE.md tech-selection rubric). Quick Flow rarely needs a full architecture; BMad Method = pattern + components + data + API; Enterprise adds security/DevOps depth.
5. Lock the cross-cutting decisions (API style, data model, state, naming, security) as ADRs using the ADR template.
6. Map every FR and NFR to a design decision in the NFR/FR coverage matrix. No orphans.
7. Fill architecture.template.md → bmad-output/architecture.md.
8. Append a one-line summary of each ADR to decision-log.md.
9. Validate (below).

Update

1. Read existing architecture.md + the changed prd.md.
2. Diff: which new/changed FR/NFR lack a decision? Which ADRs are now contradicted?
3. Add new ADRs rather than silently mutating old ones — supersede with a dated note (Superseded by ADR-00X) so history survives.
4. Re-run the coverage matrix; re-validate.

Validate

1. Run the validator on the target doc:

   bash ${CLAUDE_PLUGIN_ROOT}/skills/bmad-architecture/scripts/validate-architecture.sh bmad-output/architecture.md
   

2. Confirm: every FR/NFR appears in the matrix; every cross-cutting concern has an ADR; each ADR has Context / Decision / Consequences; trade-offs are documented.
3. Report gaps as a checklist. Do not "fix the code" — fix the plan.

ADRs — the core artifact

Each significant decision is one ADR (template: ${CLAUDE_PLUGIN_ROOT}/skills/bmad-architecture/templates/adr.template.md):

• Title — ADR-00N: <short imperative> (e.g. ADR-003: Use REST with JSON:API envelope).
• Status — Proposed / Accepted / Superseded.
• Context — the forces, including which FR/NFR drive it.
• Decision — the choice, stated so a downstream agent can follow it mechanically.
• Consequences — what becomes easy, what becomes hard, what is now LOCKED for all stories.
• Alternatives — what was rejected and why.

Minimum ADR set for a BMad-Method project: API style, data/persistence model, AuthN/AuthZ, state management, error/response convention, naming convention. These are exactly the choices that, left unstated, cause parallel agents to diverge.

FR/NFR coverage matrix

A required table in architecture.md: one row per FR and per NFR → the component(s) and ADR(s) that satisfy it → status. The validator fails if NFR mapping is missing. NFR categories and common ADR topics are in REFERENCE.md.

Validation scripts

Script	Purpose
scripts/nfr-checklist.sh	Prints the full NFR category checklist to drive systematic coverage.
scripts/validate-architecture.sh <doc>	Checks required sections, NFR coverage, ADR presence, justification, and trade-offs. Pass/fail report.

Invoke with the ${CLAUDE_PLUGIN_ROOT} prefix shown above. (The orchestrator marks scripts executable; you may also run them via bash.)

Handoff

When architecture.md validates clean, it is ready for bmad-epics-and-stories (epic sharding / story creation), then bmad-sprint-planning for wave sequencing. Stories will cite this document by section in their Dev Notes and inherit its LOCKED cross-cutting decisions, so every parallel dev agent builds against the same contract.

Reference

Detailed NFR categories, common ADR topics, and the tech-selection rubric live in REFERENCE.md.

---

Part of the BMAD Planning & Orchestrator plugin — a Claude Code harness for the BMAD Method by the BMAD Code Organization (https://github.com/bmad-code-org/BMAD-METHOD). Implements the spirit of bmad-create-architecture. All methodology credit belongs to the BMAD Code Organization.