27 conventions, hand-tuned

The philosophy

We start from the ways AI-generated code actually breaks: silent any types, swallowed errors, illegal states reachable through optional fields, runtime crashes from unvalidated input. Each rule is written to push the model away from one of those failure modes — and kept in the library only if it visibly improves the code agents produce.

Each rule is short, opinionated, and self-contained. Your agent reads one when it’s about to make a relevant change — not all of them at the start of the session. That means smaller prompts, smaller bills, and rules the model can actually keep in working memory while it writes.

What’s in the library

The current library is 27 conventions covering the areas where AI-generated TypeScript most often goes wrong: type strictness, error handling, runtime validation at trust boundaries, naming and module structure, control flow, and framework-specific patterns (e.g., React component prop shapes, Prisma model conventions).

Specific rules are available to subscribers through the MCP server — we don’t publish the full list here. Subscribe to access the corpus.

Updates

We add and refine conventions as the underlying tooling and model behavior evolve. Updates land instantly: there’s nothing to redeploy on your end. Major changes are announced to subscribers via email.