Rules file

A rules file is a document an AI coding agent reads on every task, stating how you want it to behave: your conventions, your tech stack, the patterns to follow, the things never to do. Tools like Claude load a project file (often CLAUDE.md) automatically, so the agent starts every session already knowing your standards instead of guessing.

This is one of the cheapest, highest-return things a founder building with AI can set up. Without it, you repeat the same corrections endlessly: use British spelling, never fork a component, always write a spec first. The agent forgets the moment a new session starts. Codify those once and they hold across every task.

The rules pay off most when they encode the parts of your stack the agent keeps getting wrong. Say you're storing data in Supabase and the agent keeps writing raw SQL that bypasses your row-level security; a line in the rules file ("every tenant table needs RLS plus a grant in the same migration") stops that class of mistake for good. Say your team ships through Linear and you want every change tied to an issue; state that once and the agent stops inventing its own workflow. Building UI off a Figma file? Tell it to resolve every colour and spacing to your design tokens rather than hardcoding hex values.

The discipline is to keep it lean and abstract: state the principles the agent should reason from, not a sprawling pile of one-off patches. A tight rules file is the difference between an agent that fits your codebase and one that fights it.