AI coding policy

Texcraft's policy around AI usage depends on which part of the codebase is being touched.

The core Texcraft Rust libraries are the library crates under crates/ that reimplement Knuth's TeX. Changes to non-testing code here are made by hand without AI. There are a few reasons for this:

  • One of the main value propositions of the Texcraft project is that it provides well-designed modular APIs for the internal components of TeX. We think this is currently a task that requires a lot of quality human judgment.

  • We have the highest possible code quality standards for this code, and in our experience the AI cannot (yet?) achieve them.

  • Specifically for the last point: there is a tension between writing the code in a way very similar to Knuth's TeX (very few data structures, no abstractions, highly procedural) and writing it in a more modern way. The former is better for correctness; the latter better for maintainability and re-usability. We think this decision requires quality human judgment.

  • The whole point of the project is to have fun reimplementing TeX's algorithms and getting an AI to do it sort of defeats the point. :)

There are a couple of small exceptions to this policy for these libraries:

  • Any kind of mechanical refactoring is of course okay to do with AI.

  • Claude Code is great at RCAing bugs, in part because it has access to Knuth's original source code and can see where the difference is. It will generally suggest a fix to the Rust code. In our experience this fix fails our code quality and/or architecture standards. We do use the AI-proposed fix as a guide to understanding the bug and how to create a better quality human fix.

All other Rust code under crates/ (testing and binaries) can be written with arbitrary AI assistance, but must be human reviewed and human readable. We do not want this code to devolve into some AI slop state such that it can't be worked on by humans any more.

All of the code under sites/ (Rust, Javascript, HTML, etc.) is maintained exclusively by AI and there is no expectation that the code be maintainable by humans.