Spec-driven, not vibe-coded: how we use AI coding tools
Md Nazmul Hosan ·
AI writes code faster. It does not decide what to build or whether it works. Our rules for using it on client projects.
Short answer: every change starts from a written user story, AI features start from an agreed evaluation set, and a person reviews all code before it ships.
Every engineer at Door Soft uses AI coding assistants. They are good at boilerplate, tests and refactoring, and they make us faster. They are also good at producing plausible code that does the wrong thing. Vibe coding is fine for a weekend prototype. It is not how we build software clients run their business on.
Our rules
- The spec comes first. Work starts as a user story with acceptance criteria, approved by the client.
- For AI features, the evaluation set comes first too. We agree 50 to 300 test cases the feature must pass.
- A person reviews every change. No generated code reaches production without review against the story and tests.
- Client code stays private. We do not paste client code or data into tools that train on it.
- Generated code meets the same bar. Same tests, same standards, same review.
Why this matters more for AI features
Ordinary code is usually right or wrong. AI features are right most of the time, which is harder to test. Without a written definition of correct and a test set, nobody can say whether a change made the product better or worse.
What clients see
A written spec before we start, a demo at the end of every sprint, an evaluation report before launch and a monthly report after it. See how we work.
Share
Nazmul founded Door Soft in Dhaka and leads system design and delivery. He writes about building AI that works in production for mid-size companies, and about the products Door Soft runs itself.
About the author