AI-Driven Development
A quality methodology for AI-assisted software development.
AI-assisted development delivers features at a rate that outpaces human comprehension. You approve code you did not write, the tests pass, the endpoints return 200. But your understanding of the system's behavior, its edge cases, and its cross-module interactions degrades with every feature shipped. The system evolves faster than your mental model of it.
This gap is invisible until something breaks. A handler loads the same entity three times because nobody checked the execution trace. An async consumer silently fails while the endpoint returns success. The API schema marks optional fields as required. Unit tests do not catch these. Code review does not catch these. Only exercising the system end to end and inspecting the distributed trace reveals them.
Traditional quality practices (TDD, code review, documentation) were designed for human-paced development. They are necessary but insufficient. This methodology adds six layers that address what they miss, adopted incrementally, each motivated by the failure of the previous one.
| Layer | What | Gap it fills |
|---|---|---|
| 0 | Code | The foundation |
| 1 | Observability | "I can't see what happens inside" |
| 2 | Documentation | "Nobody knows the philosophy" |
| 3 | AI Instructions | "AI doesn't follow conventions even with docs" |
| 4 | Callable Surface | "Nobody verifies the system works e2e" |
| 5 | Journey Verification | "Nobody checks quality from the user's perspective" |
Plugin
The repo includes a cross-tool plugin for Copilot and Claude Code with skills for walking journeys, evaluating criteria, designing journeys, and analyzing traces. Install instructions are in the repository README.
This is a living document. The full methodology, specs, and tooling live in the ai-driven-development repository.