Zero Drift

Zero Drift

A Claude Code skill that keeps long AI sessions grounded.
Replies open with your name. Tasks keep a TASK.md that the next session resumes from.

English · Português · Español

The problem · Quick start · The two rules · How it works · Install · FAQ

Zero Drift is a Claude Code skill with two rules, applied from the first response to the last. Install it as a plugin and it loads at the start of each session.

The Problem

Long sessions break down. The context window fills up, so you open a fresh Claude instance and burn ten minutes re-explaining what you were doing. By then the AI has started drifting: it invents work it never did and forgets decisions you already made.

Zero Drift gives you a way to catch the drift early and hand off cleanly.

Quick start

Open a new Claude Code window. If the first reply starts with your name, Zero Drift is active. Other install options are under Install.

The Two Rules

1. Named Response
The AI opens every reply with your name. This keeps replies personal and easy to scan in a long log.

The AI reads your name from git config user.name or your CLAUDE.md. If it finds nothing, it asks you once.

The name is your hallucination detector. When the model starts to drift, the name breaks first: it drops, changes, or reads slightly off. That is your signal the session is degrading. Open a new window, say *"read the TASK.md and continue"*, and resume from a clean state.

2. Living Task Document
Each task gets a TASK.md in the project root. After every meaningful prompt the AI records what it did, what broke, what it fixed, and where things stand now.

The Log is a record, not a diary. The AI cannot assert "fixed X". It shows proof: the command run, its output, the exit code. No proof, no Log entry; unproven work goes to Unverified…