Drift

Drift is what happens when an AI coding assistant gradually loses alignment with your project’s goals, conventions, and decisions over the course of a long conversation or across multiple sessions. As earlier context falls outside the AI’s memory window, it starts generating suggestions that conflict with what was already agreed on. The standard fix is persistent context files that re-anchor the AI at the start of every session.

The Simple Version

Imagine hiring a consultant who’s brilliant in the first meeting but gradually forgets what your company does over the next few weeks. By month two, they’re suggesting things that have nothing to do with your business. That’s drift.

With AI coding tools, drift happens because the AI can only hold so much conversation in memory at once (the context window). As conversations get longer or sessions reset, earlier decisions and constraints can fall out of memory, and the AI starts making suggestions that conflict with what was already decided.

Why It Matters

Drift is one of the most common frustrations when building with AI. You spend an hour getting the AI to understand your project, then the next session it acts like it’s never seen it before. Without a strategy to prevent drift, you end up repeating yourself constantly or, worse, accepting suggestions that undo good work.

How It’s Used on This Site

This site prevents drift using persistent files. PLAN.md records the project’s goals, architecture, and current status. CLAUDE.md contains operating instructions and conventions. Claude Code reads these files at the start of every session, so it picks up exactly where things left off, even when the conversation history resets.