[Workshop]
How Spec-Driven Development Changes the Way Teams Build Software
Shibi Ramachandran - ING / Ram Mohan Rao Chukka - JFrog
AI‑assisted development like Copilot can produce 50 lines of code in seconds. But that speed creates a subtle trap where we slide from engineering into vibe coding. The AI writes something plausible, we accept it, and slowly our system drifts. The Architectural intent disappears into a chat history, key decisions end up in someone’s head instead of in the codebase. We save minutes writing code, and lose hours debugging code we never really designed.
In this session, we’ll show a practical approach for taking the wheel back using Spec‑Driven Development (SDD) — a workflow my team has implemented over the past year to keep AI productive without letting it degrade architecture, consistency, or maintainability.
Rather than fighting the AI, we constrain it. SDD provides a structured, repeatable way to turn “AI vibes” into engineering discipline:
- Architect the Constitution
We define a durable set of semantic guardrails - architectural rules, invariants, banned patterns, and required behaviors. These aren’t prompts. They’re a permanent contract the AI must obey. We’ve used this to enforce logging policies, domain rules, error‑handling structures, and even naming conventions.
- Anchor the Specification
Instead of ephemeral prompts, we maintain a living, versioned specification that becomes the source of truth. All generated code must explicitly reference this spec. This prevents architectural drift and ensures new contributors (human or AI) understand the intent behind every module.
- Shift Validation Left
We validate AI‑generated code by automatically checking it against the spec and its guardrails. This catches drift instantly before review, before merge, before the code becomes a future debugging headache.
What attendees will learn:
- How “vibe coding” quietly creates systemic architecture decay
- How to define a specification that survives beyond individual prompts
- How to encode design intent into enforceable contracts
- How to build a validation loop that keeps AI output aligned with architecture
- How to adopt SDD incrementally — even in existing systems
Who this session is for:
Engineers, tech leads, and architects who want AI to accelerate them without sacrificing consistency, design quality, or maintainability.