2026년 최고의 명세서 주도 개발 AI 툴 9선

Editors Pick Agentic AI Software Engineering Staff Top As AI coding agents grow more capable, a structural problem has emerged: speed without clarity. Developers generate working code in minutes, only to discover days later that it doesn't match what the system actually needed. Spec-driven development (SDD) addresses this directly — by treating a structured specification as the source of truth and code as its generated output, rather than the other way around. This list covers the 9 AI tools that developers are actually using to implement SDD workflows in 2026. AWS Kiro 🔗 kiro.dev | Docs | Models Kiro is an agentic IDE built around spec-driven development, designed to take developers from concept to production with structured rigor instead of iterative prompting. Rather than writing code and asking an AI to help along the way, Kiro requires developers to formalize intent first. It guides them through a three-phase process — Requirements, Design, and Tasks — producing three structured artifacts: requirements.md, design.md, and tasks.md. A notable technical detail: Kiro generates user stories using EARS (Easy Approach to Requirements Syntax) notation, which produces structured acceptance criteria covering edge cases that developers would otherwise handle manually. A major differentiator is its agent hooks system — event-driven automations that fire when files are saved or created, handling tasks like test updates, README refreshes, and security scans without manual prompting. For model selection, Kiro's default is an Auto router that combines multiple frontier models — including Claude Sonnet, Qwen, DeepSeek, GLM, and MiniMax — and selects the optimal model per task to balance quality and cost. Developers can also pin a specific model for consistent behavior. Built on Code OSS, VS Code users will feel at home immediately. Kiro also supports a CLI and a web interface, and does not require an AWS account to use. Best for teams that need formal spec workflows in a familiar development environment. GitHub Spec Kit 🔗 github.com/github/spec-kit | Blog Post GitHub Spec Kit is the most community-adopted open-source option for spec-driven development — a Python CLI with 93,000+ stars, the latest release being v0.8.7 (May 7, 2026), supporting 30+ AI coding agents including Claude Code, GitHub Copilot, Amazon Q, and Gemini CLI. The workflow runs through four phases with clear checkpoints: Specify (captures business context and success criteria), Plan (translates specs into architectural decisions), Tasks (decomposes plans into testable, reviewable units), and Implement (runs AI agents under those constraints). At the foundation of every Spec Kit workflow is a "constitution" — a markdown rules file containing high-level immutable principles that apply to every change across every session. This becomes the persistent contract between the developer and the agent. Spec Kit's philosophy, as GitHub framed it, is that code is now the last-mile output: intent is the source of truth, and specifications are executable. It's the default starting point for teams new to SDD and the most portable option for teams that want to keep their existing IDE. BMAD-METHOD 🔗 github.com/bmad-code-org/BMAD-METHOD | Docs BMAD-METHOD (Build More Architect Dreams) is an MIT-licensed open-source framework that orchestrates 12+ specialized AI agents across the full software development lifecycle. Version 6.6.0 shipped on April 29, 2026, with the project reaching 46,700+ GitHub stars and more than 5,500 forks. The 12+ agents cover distinct SDLC roles — including product management, architecture, UX, development, QA, and scrum master functions — and work together through structured, file-based handoffs: each agent reads the previous agent's output document and writes its own, maintaining a traceable chain from requirements through delivery. V6 introduced the Cross Platform Agent Team, allowing the same agent configuration to operate across Claude Code, Cursor, Codex, and other hosts without reconfiguration. The V6 architecture also separates concerns into three layers: BMad Core (the universal human-AI collaboration framework), BMad Method (the agile development module built on Core), and BMad Builder (which lets teams create and share custom agents and workflows). BMAD is the go-to framework for teams that want highly structured, role-separated multi-agent workflows without vendor lock-in. The framework is entirely free with no paywalls. Augment Code 🔗 augmentcode.com | SDD Guide Augment Code approaches spec-driven development from the context layer rather than the spec authoring layer. Its Context Engine maintains a persistent architectural understanding across 400,000+ files — addressing the cross-repository context gap that breaks most specification workflows at scale, particularly in multi-service brownfield codebases. Augment reports 70.6% on SWE-bench (compared to a 54% industry average) and a 59% F-score on an AI code review benchmark; these figures are vendor-reported and should be treated accordingly. Its BYOA (Bring Your Own Agent) model lets teams plug in Claude Code, Codex, or OpenCode alongside its native Auggie agent. Augment Code does not author specs natively — teams still need a tool like Spec Kit or Kiro for structured spec management — but it provides the semantic foundation that makes those specs accurate across large codebases. Best suited for enterprise teams running complex multi-service architectures where context drift, not spec creation, is the primary failure mode. Claude Code 🔗 claude.ai/code | Docs Claude Code is Anthropic's agentic command-line tool, and unlike tools such as Cursor or GitHub Copilot that augment a developer's workflow, it is designed for fully autonomous development — planning, orchestrating multi-step workflows, and asking follow-up questions without constant prompting. For spec-driven workflows, Claude Code handles large specification documents well within a single session, processing complete requirement sets and generating implementations in one coherent pass. Developers typically use CLAUDE.md files as the spec layer — a lightweight approach that enforces persistent project context, coding standards, and architectural constraints across every session. This means many developers are already practicing a form of SDD with Claude Code without formally labeling it as such. Claude Code also serves as a commonly supported execution agent across SDD frameworks including BMAD, GSD, and GitHub Spec Kit. GSD (Get Shit Done) 🔗 github.com/gsd-build/get-shit-done GSD is a spec-driven meta-prompting and context engineering framework built primarily for Claude Code and compatible agents, positioning itself as the lean, low-ceremony alternative to BMAD. The project has crossed 61,000 GitHub stars — growing from zero to that figure in under five months since its December 2025 initial commit. It installs via npx get-shit-done-cc@latest and works across Claude Code, OpenCode, Gemini CLI, Codex, Copilot, Cursor, Windsurf, Augment, and Cline. Its multi-agent orchestration spawns parallel researchers, planners, executors, and verifiers, each operating in a fresh context window with up to 200K tokens dedicated to implementation. The model-agnostic design — including support for OpenRouter and local models — decouples the workflow from any single LLM vendor. Where BMAD adds sprint ceremonies and stakeholder coordination, GSD's philosophy is that complexity should live in the system, not the workflow. It also fills a gap that Claude Code itself doesn't cover natively: context rotation, quality gates, and planning state persistence across sessions. Cursor (with Plan Mode + Project Rules) 🔗 cursor.com | Agent Best Practices Cursor remains one of the most widely used AI editors, and its Plan Mode makes it a practical entry point for teams adopting spec-first habits without switching toolchains. Plan Mode creates a detailed implementation plan before any code is wr