How To Create Portable AI Skills Across Harnesses (Claude Code, Codex, Open Code)

Agentic engineering is the future. But it's still unclear which harness will win, if any.

I've been hedging my bets by moving to a harness-agnostic system for AI skills so no matter the harness I use, I can still leverage the skills I've developed.

Skill layout

The basic approach is:

• Core - single source of truth for instructions and skills, stored in markdown files with internal links.
• Shell - harness-specific files that point to the Core.

This allows one source of truth for your skills, with first-class support by any harness so no matter which one wins (or you prefer) you can take your instructions with you.

Example

.
├── AGENTS.md
├── agents/skills/
│   ├── code-review/SKILL.md
│   └── debug-rust/SKILL.md
├── CLAUDE.md            -> @AGENTS.md (import)
└── .claude/
    ├── commands/
    └── skills/          -> ../agents/skills/ (symlink)

Core:

• AGENTS.md - AGENTS.md is the cross-tool open standard, governed by the Linux Foundation's Agentic AI Foundation as of 2025.
• agents/skills/ - each skill has a folder with a SKILL.md describing it and whatever other files it wants. Follows Anthropic's convention: SKILL.md (required) plus optional folders for supporting artifacts like references/, scripts/, templates/, assets/.

Shell:

• Claude Code:
  • CLAUDE.md - @AGENTS.md at the top
  • Commands - for any skills I use a lot, points to them in its folder
  • Skills - symlinked to agents/skills/. Symlinks are fragile on Windows but work fine on Linux, macOS, and WSL.
• Codex / OpenCode / Cursor / GitHub Copilot / Windsurf - read AGENTS.md by default.

Next

I'm pretty bullish on agentic engineering though I'm not convinced any of the current tools will be the ones that win. We're still super early so I think it pays to find approaches that will likely be useful regardless of pending improvements.

If you're curious about how I'm doing agentic engineering, I've been taking snapshots of my ai-dotfiles repo and putting them in the HAMY LABS Code Examples Repo available to HAMINIONS members.

If you liked this post you might also like:

• 5 AI Coding Best Practices from a Google AI Director (That Actually Work)
• How I Actually Code with AI as a Senior Software Engineer
• I Built an AI Orchestrator and Ran It Overnight - Here's What Happened