One thing I’ve noticed while building AI developer tools:
AI often makes decisions that are technically correct—but project-wise completely wrong.
Examples:
- Removing “unused” code that’s intentionally reserved.
- Replacing an architecture because it looks simpler.
- Changing interfaces without knowing external dependencies.
- Refactoring away workarounds that exist for real production issues.
None of these are reasoning failures.
They’re information failures.
The AI only sees the code.
It doesn’t see the project’s accumulated knowledge.
That’s why I’ve been building Contorium, a Project Intelligence Layer.
Instead of treating a repository as only files, Contorium treats it as a collection of evolving decisions.
When AI works on a task, it can understand:
- Why a decision was made.
- Which assumptions still matter.
- What alternatives were rejected.
- How the project evolved over time.
Code tells you how something works.
Project Intelligence tells AI whether it should change it at all.
I believe the next generation of AI development tools won’t simply generate better code.
They’ll make better project decisions.
https://www.contorium.dev/
https://github.com/ContoriumLabs/contorium
Top comments (0)