What is GitHub's spec-kit?

spec-kit is an open-source toolkit from GitHub that introduces Spec-Driven Development — a workflow where you describe what you want in natural language, the AI writes a formal spec, generates an implementation plan, breaks it into tasks, and builds it. It works via slash commands (/speckit.specify, /speckit.plan, /speckit.tasks, /speckit.implement) in any supported AI coding agent.

What is Spec-Driven Development and how is it different from vibe coding?

Vibe coding means prompting an AI and hoping for the best — no spec, no plan, unpredictable results. Spec-Driven Development makes the specification the source of truth: you describe what to build (not how), the spec captures requirements precisely, and the AI implements against that spec rather than guessing your intent. The spec persists in your repo, stays current, and guides future changes.

Which AI agents does spec-kit support?

Over 25 agents: Claude Code, GitHub Copilot, Cursor, Windsurf, Codex CLI, Gemini CLI, Qwen Code, Pi Coding Agent, Roo Code, Kilo Code, Amp, Auggie CLI, opencode, JetBrains Junie, Tabnine CLI, Mistral Vibe, IBM Bob, Jules, Kiro CLI, Trae, Antigravity, and more. It also supports a Generic mode for any unsupported agent.

What is the /speckit.constitution command?

The constitution is your project's governing document — coding standards, testing requirements, performance expectations, architectural principles. It's created once and referenced by all subsequent spec-kit commands to ensure consistency. Think of it as the ground rules the AI must follow for every implementation in your project.

Does spec-kit work with existing codebases or only greenfield projects?

Both. Greenfield: initialize in an empty directory and spec-kit scaffolds the entire project from your description. Brownfield: initialize in an existing project and spec-kit adds specs for new features without requiring prior specs. Community walkthroughs exist for a 307,000-line .NET CMS and a 420,000-line Java runtime — both extended successfully using spec-kit.

How does spec-kit handle large projects and distributed teams?

spec-kit supports multi-pass implementation for large features, incremental spec updates, and branch naming strategies for teams (sequential: 001, 002, 003 or timestamp: YYYYMMDD-HHMMSS to avoid conflicts). For enterprise/air-gapped environments, it ships an offline installation path using pip download wheel bundles.

GitHub's spec-kit: Spec-Driven Development for the AI Coding Age

By Prahlad Menon Published 2026-03-22 4 min read

GitHub just released spec-kit — an open-source toolkit for Spec-Driven Development — and it crossed 79,000 GitHub stars in days.

The fact that it’s from GitHub itself matters. This isn’t a third-party developer workflow tool. It’s GitHub’s public thesis on how AI-assisted software development should work.

The Problem It’s Solving

Vibe coding works until it doesn’t. Describe something to Claude Code or Copilot, it builds something, you ask for changes, it breaks something else. After a few iterations the context is polluted, the AI has forgotten earlier decisions, and the output is a pile of code that mostly does what you wanted.

The root issue: there’s no specification. The AI is guessing your intent from conversation history, not implementing against a documented requirement.

spec-kit’s answer: make the spec executable. You write down what you want clearly once. The spec persists in your repo. Every implementation step works from it. Future changes update the spec first, then the code.

The Five-Command Workflow

# Install
uv tool install specify-cli --from git+https://github.com/github/spec-kit.git

1. Set project principles

/speckit.constitution Create principles focused on code quality, 
testing standards, and performance requirements

The constitution is created once — it’s your project’s ground rules. The AI references it for every subsequent command.

2. Describe what you want to build

/speckit.specify Build an app that organizes photos into albums 
grouped by date. Albums support drag-and-drop reordering. 
Photos display in a tile grid within each album.

Focus on what and why, not the tech stack. The AI writes the formal spec.

3. Choose your stack

/speckit.plan Use Vite with vanilla HTML, CSS, and JavaScript. 
Images stay local. Metadata in SQLite.

Now the AI has the requirements (spec) and the constraints (plan). It generates a technical implementation plan.

4. Break into tasks

/speckit.tasks

The plan becomes a structured task list. This is what the AI will execute — concrete, ordered, testable steps.

5. Build it

/speckit.implement

That’s the full loop: constitution → specify → plan → tasks → implement. The spec and plan persist in your repo as living documentation that stays in sync with the code.

Agent Support

The agent compatibility table is striking — spec-kit works with essentially every major AI coding agent:

Claude Code, GitHub Copilot, Cursor, Windsurf, Codex CLI, Gemini CLI, Qwen Code, Roo Code, Kilo Code, Amp, opencode, JetBrains Junie, Pi, Tabnine, Mistral Vibe, IBM Bob, Jules, Kiro, Trae, Antigravity, and more.

The --ai generic flag lets you point it at any agent not on the list. It generates the right format (slash commands, skills files, or custom prompt templates) for each agent’s interface.

Works on Existing Codebases

The greenfield demo (describe an idea → build it from scratch) is the obvious use case. But the brownfield support is more interesting for most developers.

Community walkthroughs include extending a 307,000-line .NET CMS and a 420,000-line Java Maven project — both with no prior specs in the codebase. spec-kit generates a constitution for the existing codebase, writes specs for the new feature, and implements against them without touching unrelated code.

This is the real-world case: you have a live system, you want to add a feature cleanly, and you want the AI to work from a spec rather than reading 300K lines of context.

The Broader Point

The tools that have accumulated the most GitHub stars in 2026 — spec-kit, Understand-Anything, SocratiCode, acpx — all solve the same underlying problem from different angles: AI coding agents need structure to be reliable at scale.

Understand-Anything gives agents a knowledge graph so they understand the codebase. SocratiCode gives agents semantic search so they retrieve relevant context efficiently. acpx gives agents a structured protocol instead of PTY scraping. spec-kit gives agents a spec so they implement predictably instead of guessing.

The context rot problem — where AI agents degrade as the context fills — is ultimately a structure problem. Vibe coding provides no structure; spec-driven development provides a lot of it.

GitHub building this and open-sourcing it is a signal about where AI-assisted development is going. The era of “describe your idea and hope for the best” is short. The era of “describe your idea, capture it precisely, and implement against the spec” is what comes next.

→ github.com/github/spec-kit