Jason Goecke, Managing Partner, Cultiv8 Ventures LLC

Following in our series of AI Coding.

Introduction

AI Without Specifications = Expensive, Fast Chaos.

The reality: Vague Idea → AI Interprets → Wrong Code Generated Fast → “That’s not what I meant” → More AI → More Wrong Code → Faster Failure Cycles → Expensive Chaos. AI doesn’t fix unclear requirements. It amplifies them at 10x speed.

What Is Spec-Driven Development? A development method where executable specifications—not verbal requirements or user stories—serve as the single source of truth defining WHAT to build, WHY to build it, and HOW to prove it’s correct.

SDD creates the precision foundation that enables AI to deliver production-ready code systematically, transforming development from an unpredictable craft into a scalable engineering discipline.

It MUST be:

• Performance requirements (measurable)
• Security requirements (testable)
• Reliability requirements (verifiable)
• Compliance requirements (auditable)

GitHub Spec-Kit is the framework that makes Spec-Driven Development practical and scalable. Built by GitHub Engineering for GitHub.com’s 200M+ users, it’s the foundation that transforms SDD from theory into competitive reality. This isn’t experimental technology—it’s production-proven infrastructure available to your organization today.

Jason’s Presentation and Discussion

This is a great review of spec-driven development, Jason keeps the presentation to the point, only 7 minutes, drawing on his vast development experience. Simply, this is the new development process all software development groups should follow. You’ll see everyone in the session agrees with Jason. So let’s review his journey to SDD.

Jason’s Journey

AI coding assistants offered a powerful “dopamine hit,” making prototyping effortless. But this immediate gratification often masked technical debt, generating “feels right” code optimized for impressiveness, not long-term team scalability. Code used for a prototype is not the same as code used for production.

Jason took his usual approach in improving software development processes. The core idea is context engineering transforms AI from a basic assistant into a professional by providing the Quality Management System (QMS) it inherently lacks. Covering sprint definition, unit test coverage, etc.

Here Jason used Roo Code, an open-source, AI-powered autonomous coding agent available as an extension for Visual Studio Code (VS Code), Rob Pickerings favorite tool. It integrates large language models (LLMs) directly into your editor to help with tasks like generating code, debugging, refactoring, and writing documentation across multiple files.

And Roo Commander is a collection of custom modes that uses Roo Code to orchestrate complex software development workflows using a multi-agent system. Jason set quality standards for how each agent did tasks such as linting, the automated process of analyzing source code to identify programmatic errors, bugs, stylistic issues, and suspicious constructs.

The gap was codifying the processes of agents.

Stop treating AI like a fast typist. Treat it as a specification execution engine, there was broad agreement on that point. Context engineering defines standards and patterns (the how), while specifications capture explicit intent (the what and why). Together, they create exponential efficiency.

The Key Point

Context engineering and explicit specifications make implementation nearly deterministic and verifiable, creating a system where AI truly executes human intent.

Jason uses Spec Kit from Github.

The Spec Kit bridges the gap between human intent and AI execution, moving away from vague “vibe coding”. Its key artifacts include:

• constitution.md: Outlines the project’s non-negotiable principles, such as coding standards, testing requirements, and architecture constraints.
• spec.md: A detailed description of the project’s goals, user journeys, and requirements (the “what” and “why”).
• plan.md: Describes the technical approach, architecture, and dependencies (the “how”).
• tasks/: A directory containing individual, actionable work units derived from the plan for the AI to execute.

What is missing is the linkage to JIRA or Github issues. But Jason counters the software development process, in the age of AI, has become different.

Jason frankly ends on, adoption of SDD is not an option, its critical to an organization’s survival. There is broad agreement from the audience, most are also using SDD.

Discussion

During Jason’s presentation, Thomas took the recent IETF VCON zip spec, vCon Zip Bundle (.vconz) file format for packaging one or more vCon conversation data containers with their associated media files into a single, self-contained ZIP archive. He popped it into Cursor, and now has a Python implementation that passed all the tests. That is power of SDD in practice.

Thomas recommend using AI to check the specs, and make sure they are complete. Joao uses AI for analysing SIP traces, i.e. messages exchanged during a SIP call to help diagnose and troubleshoot issues in a VoIP.

Joao also mentioned the recent launch of Google Antigravity, an agentic development platform, evolving the IDE into the agent-first era. He hinted at the trend to make it easier for non-coders to use. Definitely today the tools are built for those with coding experience, but the use of natural language is making the addressable market broader.

Jason mentioned his work with Pulumi, Next-Level Infrastructure as Code, with Agentic AI. Any cloud, any language. Joao highlighted marketing and sale folks are building tools, and the quality of the code generated is important, as they can specify BUT do not not have the skills to make a quality assessment. Hence the importance of SDD.

Jason highlighted the importance of the constitution in Spec Kit, and how it should be used in organizations where compliance matters, because of the radical transparency in SDD, certification could be automatic!

Rob raised a question on architecture, in Spec Kit plan.md: describes the technical approach, architecture, and dependencies (the “how”). And it comes after the business specific, so it subordinate to the business needs.

Product managers will becomes the elite coder. In time the expertise of the experience developers become baked into SDD. So the product manager or junior develop are at less of an disadvantage. They’ve grown up with the tools, so are skilled in its application to solving business problems. Rob also highlight the tools themselves are great for learning.

Jason wraps up asking if anyone uses tools other than Spec Kit. Thomas highlight he adds an AGENTS.md, a standardized markdown file used in software projects to provide specific context and instructions to AI coding agents. It functions as a “README for machines.”

Thank you to everyone for a powerful session, it’s time to move to Spec-Driven Development.