Eliminating AI Agent Instruction Drift with agent-kit
These articles are AI-generated summaries. Please check the original sources for full details.
One AGENTS.md for Every Coding Agent: Auto-Derive CLAUDE.md, GEMINI.md & Copilot Instructions
Hasan Zohdy introduced @mongez/agent-kit to solve the fragmentation of AI coding tool instructions. The system automatically synchronizes a single source of truth across multiple proprietary agent configuration paths.
Why This Matters
Different AI agents require instructions in different files (e.g., .github/copilot-instructions.md vs .gemini/GEMINI.md), leading to manual duplication and ‘drift’ where agents disagree on project conventions. This fragmentation forces developers to maintain near-identical copies of documentation, increasing the risk of inconsistent agent behavior across a single codebase.
Key Insights
- The emerging open standard for project instructions is AGENTS.md, natively read by Codex, Cursor, Amp, OpenCode, and Goose (2026).
- Instruction derivation allows one file to populate CLAUDE.md, .gemini/GEMINI.md, .github/copilot-instructions.md, and CONVENTIONS.md simultaneously.
- @mongez/agent-kit enables npm packages to ship ‘skills’ (SKILL.md files) that are automatically mirrored into agent directories like .claude/skills/ upon installation.
- Automated flattening solves Claude Code’s limitation where it only discovers skills at the top level of .claude/skills/, converting nested paths like backend/auth into flat names like backend-auth.
Working Examples
Installation and initialization of the agent-kit CLI.
npm install -D @mongez/agent-kit
npx agent-kit init
Adding agent-kit sync to package scripts to ensure instructions stay fresh after every install.
{
"scripts": {
"postinstall": "agent-kit sync"
}
}
Targeting specific agents for skills export or running derive-only mode for Gemini and Aider.
npx agent-kit sync --target claude,cursor,codex
npx agent-kit sync --derive-only
Practical Applications
-
Use case: Library authors bundling SKILL.md files in npm packages so that any developer installing the dependency automatically grants their coding agent knowledge of how to use the library.
Pitfall: Manual copy-pasting of skills into .claude/skills/, which leads to folder collisions and maintenance overhead as skill sets grow.
References:
- [type: text, text: ‘I would like you to help me draft an email that expresses my interest in a particular job vacancy but highlights that I am currently working in another role and wish to keep my application confidential.’}]
Continue reading
Next article
Anthropic Releases Claude Opus 4.8: #1 on Benchmarks, Parallel Subagents, and It Actually Tells You When Your Code Is Wrong
Related Content
Stop Sharing Prompts: Scaling AI Productivity with Claude Plugin Marketplaces
Transition from fragmented local prompts to a unified Git-based marketplace for Claude plugins to ensure team-wide tool synchronization and version control.
AI Coding Assistant Comparison 2026: Cursor, Copilot, Claude Code, and JetBrains AI
A technical evaluation of 2026's AI coding tools, where Cursor leads power users with a 200K context window and agentic refactoring.
Dinghy: Unifying DevOps Tooling with a Single CLI and Docker Engine
Dinghy unifies infrastructure, diagrams, and docs into one CLI, allowing engineers to generate 248 lines of Terraform from just 8 lines of TSX source.