Spec Kit vs OpenSpec vs BMad: Choosing a Spec-Driven Development Framework

The spec-driven coding ecosystem sorted itself into three big philosophies by mid-2026. GitHub Spec Kit (~115k stars) is the standardizing workflow: a chain of versioned artifacts enforced by a constitution. OpenSpec (~55k stars) is the fluid engine: four commands, delta specs, minimal ceremony. BMad Method (~50k stars) is the virtual team: a cast of specialist agents running the whole software lifecycle with adversarial review.

They all implement the same core idea — write the spec first, then let the agent build — but they answer the three questions that matter differently:

What is the spec? A versioned source of truth? A work item in a pipeline? One artifact among many?
Who is responsible for quality? The artifact chain? The process gates? The team structure?
What does adoption cost? The answer that decides whether you'll actually use it.

The head-to-head

| | GitHub Spec Kit | OpenSpec | BMad Method | |---|---|---|---| | Philosophy | Code serves the spec — spec is truth | Fluid, iterative, change-based | Full-lifecycle virtual agile team | | Workflow | Constitution → specify → clarify → plan → checklist → tasks → analyze → implement → converge | Propose → explore → apply → archive | Agentic planning (Analyst→PM→UX→Architect) → development (Scrum Master→Dev→QA→review) | | Spec format | Full spec files per feature branch, numbered requirements | Delta specs (ADDED/MODIFIED/REMOVED) per change | PRD with MoSCoW + Gherkin, UX spec, architecture | | Governance | Constitution, semantically versioned; compliance gates | None — discipline is yours | Formal phase gates + Implementation Readiness Check | | Review | Human reviews the PR (spec travels with it) | None built in | Built-in adversarial review (zero findings triggers re-analysis) | | Parallelism | Branch-per-feature, [P] task markers | Change-per-feature, conflict detection at proposal | Branch-per-story, party mode | | Setup | uvx/pipx, ~30 min incl. constitution | npm, ~5 min | npx bmad-method install, 1–2 hours to learn | | Learning curve | Moderate | Low | Steep (agile vocabulary, many roles) | | Stars (mid-2026) | ~115k | ~55k | ~50k | | Best for | Teams standardizing on one workflow | Brownfield speed; individuals | New products; full traceability; regulated work |

Where each one wins

GitHub Spec Kit wins when the problem is drift

If your failure mode is "the agent built something adjacent to what I asked," Spec Kit is the strongest answer: the spec is a reviewable, mergeable artifact with numbered requirements, ambiguity is flagged ([NEEDS CLARIFICATION]) rather than guessed, and the constitution is enforced at plan time. It's the safest default for a team that wants one standard workflow across Claude Code, Copilot, Cursor, and whatever the team buys next year. The cost: ceremony. Small changes still generate spec/plan/tasks artifacts, and the full 9-command path is heavy for experiments — use the lean path.

OpenSpec wins when the problem is adoption

If your team won't adopt a process that feels like paperwork, OpenSpec removes every excuse: four commands, no constitution, no gates, ~5-minute setup on an existing codebase. Delta specs are a genuine innovation — reviewers see what changed in the requirements before code exists. The cost: nothing enforces anything. No adversarial review, no gates, no memory between sessions. You get exactly the discipline you bring. If you choose OpenSpec, seriously consider pairing it with execution discipline from Superpowers or a fusion project.

BMad wins when the problem is coverage

If you're building a new product and want AI engaged from the first idea to the deploy, BMad covers ground the other two don't: a real PRD with Gherkin acceptance criteria, a UX spec, architecture, QA, and — uniquely — adversarial review as a built-in role with the "zero findings is suspicious" rule. The cost: the steepest learning curve in the ecosystem and a process that genuinely can be overkill for small work (its own Quick Flow track notwithstanding). Use it when the work is big and the stakes are traceability.

The honest tiebreakers

Star counts shouldn't decide for you. The 2026 community analysis is explicit that several spec-driven repos show inflated growth — stars are a hype signal, and the most rigorous projects aren't always the most starred. The frameworks' relative engineering quality is roughly: Spec Kit is the most battle-tested ecosystem (100+ extensions), SpecD (not in this comparison, but covered here) is arguably the most rigorous engineering, and BMad is the most ambitious scope.

The real decision is failure mode. Pick the framework that targets the failure you actually have:

Agent drifts from requirements → GitHub Spec Kit
Team won't adopt ceremony → OpenSpec
Sloppy, untested execution → Superpowers (or OpenSpec + fusion)
Need full traceability on a big product → BMad
Solo, want zero setup → specdd

The scaffolding-vs-truth question underneath

All three frameworks sit on a philosophical spectrum about what a spec is: Spec Kit's "code serves the spec" (spec as truth), OpenSpec's living delta store (spec as shared memory), BMad's planning artifacts (spec as process inputs). If that argument sounds like the one you've been having in your team, the spec-driven vs vibe coding comparison frames it, and the complete guide has the full ecosystem map.