AGENTS.md: Unified Rules for Agentic Coding with Refusal Principles
A single AGENTS.md unifies Karpathy’s refusal principles and Boris Cherny’s Claude Code workflow, enforcing minimal changes, checks, and clear questions.
One more AGENTS.md that combines Karpathy’s four refusal principles in agentic programming and the Claude Code workflow by Boris Cherny.
Key changes: the agent will challenge your mistakes, make only the minimally necessary changes, won’t touch unrelated code without a reason, will run checks before reporting completion, and will ask questions whenever there is any ambiguity.
The file is about 200 lines long; only two blocks are edited (project context and accumulated project experience), the rest are fixed, concise rules of conduct. Thanks to symlinks for CLAUDE.md and GEMINI.md, a single file can centrally govern the behavior of all agents.
Comments
Weekly digest
The best of vibe coding, AI agents and open source — once a week, no spam.
Related articles
All articlesNew Claude Code Mod Adds Configurable Agent Flow Sidebar
A new Claude Code mod adds a configurable sidebar that displays all active subagents, their actions and context usage, making agent management easier.
Claude Code Desktop Boosts Productivity with Parallel Sub‑Sessions
Claude Code now runs parallel tasks in separate sub‑sessions, lets you manage environments, worktrees and launch new tasks with one click.
AntiVibe Skill for Claude Code: Turn Any Codebase into Lessons
Use AntiVibe with Claude Code to transform any codebase into detailed learning material, explaining architecture, logic, and trade-offs.