GitHub has put a clear idea on the table, one that many development teams had already started to sense: asking an AI to “build an app” and hoping for a good result is not enough when the goal is to create maintainable software. Its new open-source project, Spec Kit, points directly to that shift. The tool proposes moving from “vibe coding”, based on broad prompts and a lot of trust in the agent, to development guided by specifications that AI can read, discuss and execute.
The github/spec-kit repository has already surpassed 94,000 stars on GitHub, a figure that reflects the interest it has generated among developers, product teams and advanced users of coding assistants. The official Spec Kit documentation defines this approach as Spec-Driven Development, a methodology in which the specification stops being a preliminary document that is abandoned once programming begins and becomes an active part of the process.
From improvised prompt to executable specification
Spec Kit’s promise is easy to understand: before asking for code, you need to define clearly what you want to build. It is not about removing programming agents from the process, but about giving them better instructions, more context and clearer limits.
The basic workflow starts with a series of commands that structure the work. First, /speckit.constitution defines the project rules: code quality, testing standards, user experience consistency or performance requirements. Then, /speckit.specify describes what needs to be built and why, without yet entering into the technology stack.
That nuance matters. In many current uses of AI for programming, the user mixes functional requirements, technical preferences, design ideas and improvised changes in the same conversation. Spec Kit tries to separate those layers. First, the intention is defined; then doubts are clarified; later, the architecture is chosen.
The /speckit.clarify command introduces a step that may be more relevant than it seems: it forces the agent to ask questions when it detects ambiguity. Instead of filling in the gaps on its own, the AI returns product decisions to the human. Then /speckit.plan allows the stack to be chosen, /speckit.tasks turns the plan into tasks ordered by dependencies, and /speckit.implement builds the feature following that path.
The result is not just generated code. It is a living specification, with traceability between intention, plan, tasks and implementation. For small projects it may look like an extra step. For professional teams, it can be the difference between an impressive demo and a software base that can be reviewed, tested and maintained.
Why this approach challenges “vibe coding”
The term “vibe coding” became popular to describe a very fast way of creating software with AI: explain an idea in natural language, let the model generate code and correct it along the way. It has worked for prototypes, internal tools and personal experiments, but it has also exposed its limits.
The problem is not that AI writes code, but that it often does so without knowing the product rules, acceptance criteria, security restrictions or architecture decisions with enough precision. As the project grows, each correction can introduce new errors, duplicate logic or drift away from what the user originally intended.
Spec Kit responds to that weakness with an idea closer to classic software engineering: AI should not guess the system, it should execute a specification. The deeper shift is moving from “AI writes code” to “AI interprets and executes a structured intention”.
GitHub presents the project as an open tool for focusing on product scenarios and predictable outcomes instead of writing every piece from scratch. It also stresses that spec-driven development is not tied to any particular language, framework or provider. The documentation mentions scenarios such as greenfield development, creative exploration with several parallel implementations and improvement of existing projects.
Compatibility reinforces that idea. Spec Kit works with more than 30 programming agents, including Claude Code, GitHub Copilot, Cursor, Codex CLI, Gemini CLI, Windsurf, Qwen Code, Roo Code, Tabnine and Devin for Terminal. In some cases it uses /speckit.* style commands; in others, such as Codex CLI in skills mode, it is invoked with $speckit-* commands.
What could change for companies and developers
For an individual developer, Spec Kit can work as a way to bring discipline to AI-assisted work. It forces users to think before executing, separate requirements from technology and document decisions that would normally be lost in a chat history.
For a company, the impact may be greater. If AI agents are integrated into real development workflows, it is not enough for them to produce code quickly. They must respect internal standards, security criteria, architecture conventions and review processes. Spec Kit includes community extensions for integration with GitHub Issues, Jira, Azure DevOps, Confluence, security reviews, task validation and drift control between specification and implementation.
That does not mean the tool solves quality problems on its own. The documentation itself warns that community extensions are maintained by third parties and should be reviewed before installation. It is also worth remembering that an agent can execute a bad specification very efficiently. Responsibility for design, product priorities and final validation remains human.
Even so, GitHub’s message fits with the recent evolution of AI-assisted development. The first wave of excitement focused on code generation. The next phase seems to be moving towards more structured systems, where AI participates in definition, planning, execution and review, but within a more controlled framework.
Spec Kit does not kill vibe coding as an informal practice. It will continue to be useful for quick prototypes and proofs of concept. What it does do is draw a line: when software matters, the conversation with AI needs to become a specification. And that specification is starting to become the real deliverable.
Frequently asked questions
What is GitHub Spec Kit?
Spec Kit is an open-source GitHub toolkit for applying spec-driven development with AI agents. It helps define requirements, clarify ambiguity, plan tasks and execute implementation through a more structured workflow.
Does Spec Kit replace developers?
No. The tool helps organize work with programming agents, but product, architecture, security and validation decisions still require human judgment.
What is the difference between Spec Kit and vibe coding?
Vibe coding usually starts from broad prompts and successive corrections. Spec Kit proposes creating a clear specification first, turning it into a technical plan and generating tasks before implementation.
Which AI agents does Spec Kit work with?
The official documentation indicates support for more than 30 agents, including Claude Code, GitHub Copilot, Cursor, Codex CLI, Gemini CLI, Windsurf, Qwen Code, Roo Code and Tabnine, among others.