GitHub shifts spec‑kit to structured specs

GitHub’s Spec Kit is basically a workflow manager for AI coding agents. You do not just tell Claude or Cursor “build me X” and hope for the best. You turn the request into a spec, then a plan, then tasks, and only then let the agent start writing code. That was always the pitch. What is changing now is the emphasis — GitHub is pushing the structured parts much harder, and the repo itself now spells out that this is about predictable outcomes instead of vibe coding from scratch. ### What is Spec Kit actually? It is an open-source toolkit plus a CLI called `specify` that scaffolds a spec-driven workflow inside a project. The core artifacts are plain files — things like `spec.md`, `plan.md`, and `tasks.md` — plus agent-specific command integrations. GitHub’s own framing is that specifications stop being throwaway docs and become executable project artifacts that guide generation, validation, and implementation. (github.blog) ### Why move away from “just prompt it”? Because one-shot prompting breaks the moment a project gets real. GitHub’s writeup from September 2025 says the problem is not that coding agents cannot write code — it is that they need unambiguous instructions. A vague prompt can look fine in a demo, but once the codebase gets bigger, the agent forgets prior decisions, misses intent, or solves the wrong problem. Spec Kit is GitHub’s answer to that gap. (github.blog) ### What feels different now? The docs now explicitly recommend a structured clarification pass before planning. The README says you should run `/speckit.clarify` before creating a technical plan because it reduces downstream rework. That is the tell. GitHub is not treating clarification as a nice extra anymore — it is turning it into a first-class step in the pipeline. ### Why does that matter for Claude and Cursor? (github.blog) Agents differ in how they ingest context, expose commands, and hold onto prior decisions. Spec Kit’s integration layer is built around that reality. GitHub keeps one integration package per agent, with a registry that tracks directories, formats, capabilities, and context files. In other words, the contract lives in files and templates, not in the model’s memory. That makes behavior more portable when a team switches between Claude, Cursor, Copilot, and others. (github.com) ### Is GitHub still broadening the tool? Yes — and fast. The repo was sitting around 94,000 stars when crawled today, and the latest tagged release was v0.8.7 on May 7, 2026. Recent changes include new integrations and more governance-oriented features, like loading constitution context during implementation. That is another sign this is becoming more formal and policy-aware, not less. ### So is this anti–vibe coding? Not exactly. (github.com) GitHub still talks about starting from a high-level description. But the workflow immediately tries to pin that intent down into structured artifacts. The point is not to kill natural language. The point is to stop treating natural language as the only interface. ### What is the bigger shift here? Spec Kit is part of a wider change in AI development culture. (github.com) Early agent workflows were chat-first and memory-heavy. The newer pattern is artifact-first — write the assumptions down, make them inspectable, and let different tools consume the same contract. That lowers the odds that an app gets rejected, rewritten, or derailed because one agent interpreted a fuzzy prompt differently from another. That last part is an inference from the way GitHub has structured the project and integrations, but it fits the direction of the tool. (github.blog) ### Bottom line? GitHub is not really shipping a prettier prompt template. It is trying to turn AI coding into a spec pipeline — one where the durable thing is the structured spec, and the model is just the interchangeable execution layer. (github.com 1) (github.com 2)