Beyond vibe coding: the case for spec-driven AI development

AI is creating a legacy code mess

AI-generated code is flooding enterprise codebases, creating a future mountain of technical debt. Matthias Steiner, vice president of Global Business Innovation at Syntax, argues that today’s productivity gains from tools like AI coding assistants will become tomorrow’s maintenance nightmare without proper governance.

“Nobody gets paid to write code but to create outcomes,” Steiner says. “Coding is only part of the job.” He warns that the freewheeling approach of “vibe coding”—where developers prompt an AI to build apps—is insufficient for software meant to last decades.

The spec-driven development solution

Steiner advocates for “spec-driven development,” which applies generative AI across the entire software development lifecycle. The core of this approach is a detailed functional specification that acts as a single source of truth.

From this spec, AI agents can consistently generate designs, code, tests, and documentation. Analyst Brad Shimmin of the Futurum Group calls this “the future maturation of software development in the agentic age,” updating older concepts like literate coding for AI workflows.

Open frameworks enabling this shift include:

• SpecKit
• OpenSpec
• Claude Task Master

Governance is the critical bottleneck

The central problem is that AI’s productivity gains could backfire. Steiner points to Jevons’ paradox: as AI makes building software faster and cheaper, the total volume of software produced will explode.

“With the gain in productivity, the number of applications will grow tremendously,” he says. “And who takes care of all of that?” His answer is that AI changes the speed of construction, but the foundational need for strong architecture, modularity, and component reuse remains.

A venture capital model for building

Syntax is putting this theory into practice. Steiner’s team of 30 engineers is running 10 product builds in parallel using a venture capital-style portfolio model.

They assume half the products will fail and be abandoned, a strategy made viable by the shortened development cycles from spec-driven AI. Their first successful end-to-end build using this method is ShiftBook, a manufacturing shift handover app that integrates with SAP.

For tooling, the team primarily uses:

• Anthropic’s Claude for coding
• Task Master for spec-driven workflow management
• TypeScript as their primary programming language

Software engineering is more relevant than ever

Steiner forcefully rejects the idea that AI is making software engineering obsolete. “Don’t call software engineering dead just yet,” he says. “I think that it’s the opposite.”

He argues that while AI handles micro-decisions and code generation, the macro-decisions still require human expertise. Defining system boundaries, managing dependencies, governing patterns, and aligning technology with business outcomes are tasks that remain firmly in the human domain.