Vibe Coding: The Appeal and the Trap
"Vibe coding" is a term popularized to describe a new way of working: describe what you want, let AI generate it, run it, see if it works, iterate. Not much reading of the generated code. Not much detailed review.
The result? Incredible speed — at first.
Problems appear after 2-3 weeks. You have a codebase that works (maybe), but you don't quite understand why. Every new change has the potential to break something unexpected. Trivial bugs take hours to debug because there's no foundation to hold onto.
This is "velocity debt" — a technical version of debt that seems fine until it suddenly has to be paid all at once.
TDD as a Safety Net
Test-Driven Development in the context of AI-augmented development isn't about writing every unit test manually. It's about:
**1. Defining behavioral contracts before implementation**
Before asking AI to write code, write — or ask AI to write — tests that define what should happen. This forces you to think about edge cases and expected behavior before getting bogged down in implementation details.
**2. Automatic verification that AI output meets expectations**
When AI generates an implementation, the existing tests immediately verify if the results are correct. No need to manually "run and see" for every scenario.
**3. Confidence to refactor and iterate fast**
This is the most valuable part. With a solid test suite, you can ask AI to refactor or optimize old code without fear of breaking existing functionality. Green tests = safe to deploy.
The TDD + AI Workflow We Use
In our sprints, the process usually looks like this:
1. **Define acceptance criteria** — Write user stories and expected behavior in plain language
2. **Generate test scenarios** — Ask AI to write test cases based on acceptance criteria
3. **Review and approve tests** — Developer review: do these tests really capture what we want?
4. **Implement against tests** — Ask AI to generate an implementation that makes the tests pass
5. **Review implementation** — Developer review of the generated code, especially for security and architecture concerns
6. **Iterate** — If something is missing, add tests, iterate
Steps 3 and 5 are where developer expertise is most needed. AI is very good at generating code; humans are much better at evaluating if that code is *correct* for the specific context.
What Tests are Most Valuable?
Not all tests are equally important. In the context of AI-augmented development, our priorities are:
- **Integration tests**: Verify that components work together. These are the most likely to "find" bugs generated by AI.
- **Happy path E2E tests**: Ensure major user journeys work from end to end.
- **Boundary/edge case unit tests**: For critical business logic.
Unit tests that only test a single trivial function? Let AI generate them automatically, but don't make them a priority.
Conclusion
Vibe coding is fast. TDD is solid. The combination of both — generating fast with AI, verifying with meaningful tests — is the way of working we believe in: moving fast without breaking things.
In an era where AI can generate hundreds of lines of code in seconds, the human ability to define "correct" becomes more valuable than the ability to write code manually.
Interested in building products with an AI-augmented workflow?
We use this approach on every client project.