Most people who use AI to build software do something like this: type a rough request, look at what comes out, nudge it, repeat. It works fine for small things. But on anything bigger — a booking system, a client portal, a custom tool — the AI starts drifting. It forgets decisions it made two steps ago. It contradicts itself. The result looks like it was assembled by committee at 2am.
GitHub just open-sourced something called Spec-Kit that approaches this differently. The idea is simple: before the AI writes a single line, you write a structured brief — what the thing should do, what it should never do, how decisions get made. The AI then works from that document as its source of truth, checking back in at each stage.
Think of it like handing a contractor a proper scope of work versus texting them a rough idea. Same person, very different outcome.
It already works with most of the major AI coding tools — including Claude, Copilot, and Cursor — and there are extensions for connecting it to project management software if your team uses that.
The deeper point isn't really about code at all. It's about the fact that AI works better when it has a clear, written brief to work from. That's true whether you're building software, generating reports, or managing a content workflow.
Spec (specification): A written document that describes exactly what something should do and how. Architects have blueprints. Spec-Kit is the blueprint for AI-built software.
Open-source: The code is publicly available for anyone to use, inspect, or adapt — usually free. Not a product someone is trying to sell you.
AI agent: An AI that doesn't just answer questions — it takes a sequence of actions to complete a task on your behalf, like a very fast intern following a checklist.
Vibe-coding: The (slightly affectionate, slightly alarming) term for building software by feel — prompting an AI and improvising rather than working from a plan.
If you work with developers or an agency that uses AI tools, it's worth asking: do you work from a spec, or are you prompting as you go? The answer tells you a lot.