Instructions for AI assistants

Each module gives your assistant a single, carefully written and tested behavior. Copy one into your assistant.

Start with the story

Good instructions need evidence

Effective instructions are hard. Knowing they are effective is even harder, and most people never check.

A model can look Socratic on its first reply, then drift as a student pushes back, asks again, or goes looking for a shortcut.

This project combines instructions with evaluations that measure how well those instructions hold up in practice.

Start with the story or learn how a run works.