A software team built a Git workflow tool almost entirely with AI, discovering that success required detailed guidelines and architectural documents rather than just prompting the AI to code.

What happened: Over 13 months, two developers used Claude Code and Cursor to write approximately 29,000 lines of Go code for git-flow-next across 10 releases, producing 392 commits. They handled project architecture and specifications while the AI wrote most of the actual code, despite neither developer being proficient in Go.

Why it matters: The team found that AI-assisted development is not about giving instructions to the AI, but rather building the system that tells AI what to do—meaning that guidelines, architectural documents, test patterns, and configuration rules became more valuable than the code itself. When they tried refactoring and rewrites without this framework, the AI consistently failed; once they documented their philosophy and constraints in eight guideline documents, each new session compound the benefits and mistakes stopped repeating.

What to watch: The team identified that AI does not learn across sessions without documentation engineering, and that certain guidance documents—especially GIT_TEST_SCENARIOS.md for test setup—prevented entire classes of bugs. Model upgrades are not strictly monotonic: when Sonnet 3.7 was released, it performed worse on this codebase than its predecessor, showing that newer models do not automatically solve prior problems.