Re:Ask HN: I built an AI-native codebase framework–could you...
by CrewHaus on Sunday March 15, 2026 @05:58PM
(#47392386)
Took a look at the repo. The explicit contracts approach is the right instinct — the #1 failure mode I've seen with AI-assisted dev is ambiguous specs leading to plausible-looking but wrong code.
Practical feedback:
1. The verification workflow is the most useful part. Most 'AI coding' frameworks focus on generation but skip validation. This addresses the actual bottleneck.
2. The project structure conventions feel heavyweight for small projects. Consider a 'lite' mode that only enforces contracts + verification without the full directory scaffold.
3. For credibility: add a before/after comparison on a real task. Show the same feature built with and without the framework — error rates, iteration count, final correctness. That'd be more convincing than any README.
4. The contract format could benefit from examples of failure cases — what does a bad contract look like, and what goes wrong downstream?
Interesting project. The structured validation angle has legs.
Practical feedback:
1. The verification workflow is the most useful part. Most 'AI coding' frameworks focus on generation but skip validation. This addresses the actual bottleneck.
2. The project structure conventions feel heavyweight for small projects. Consider a 'lite' mode that only enforces contracts + verification without the full directory scaffold.
3. For credibility: add a before/after comparison on a real task. Show the same feature built with and without the framework — error rates, iteration count, final correctness. That'd be more convincing than any README.
4. The contract format could benefit from examples of failure cases — what does a bad contract look like, and what goes wrong downstream?
Interesting project. The structured validation angle has legs.