Behave — Development skill for Claude Code

CLAUDE.md That Works

[](https://zeljkoavramovic.github.io/behave/)

One file. Zero dependencies. Discipline every AI coding session.

Overview

A global `CLAUDE.md` that makes AI coding assistants actually behave, so every coding session becomes more disciplined, more predictable, and more productive.

It works natively with **Claude Code** and **OpenCode**. For other agent tools, reuse the same framework through their preferred instruction file convention, such as `AGENTS.md` or `GEMINI.md`. No setup, no dependencies, one file.

The ideas come from **[Andrej Karpathy's](https://x.com/karpathy/status/2015883857489522876)** observations on LLM coding pitfalls (**[Multica adaptation](https://github.com/multica-ai/andrej-karpathy-skills)**), and from **[David Scott Bernstein's](https://github.com/ThePassionateProgrammer/knowledge-base-starter)** partnership-driven working-agreement style. From there it grew into something more systematic, tuned for real projects.

There is also a **[single-page walkthrough](https://zeljkoavramovic.github.io/behave/)** - with visual diagrams, side navigation, before/after comparison, feature comparison table, FAQ accordion, and one-click install commands.

The problem it solves

AI coding assistants, left unchecked, tend to:

• Start coding before clarifying ambiguous requirements
• Add unrequested features, abstractions, and "improvements"
• Expand scope mid-task, refactor while fixing or "improve" while adding
• Silently rename files, change APIs, or rewrite working code
• Skip verification until after the mistake is already in the diff
• Suppress errors without telling you
• Push to version control or delete files without warning
• Run destructive CLI commands or modify secrets with no safety gate
• Ship incomplete or stub implementations and call them done
• Pick arbitrarily when rules conflict, choosing he