How to Share Skills Across Multiple AI Agents

I run Claude Code, Codex, and OpenClaw on the same machine. Every day. And every day I hit the same wall: switching between agents means starting from scratch. They don’t share context. They don’t share state. You end up copy-pasting your way through conversations just to get the next agent up to speed.

The Skills approach changed that for me. Not perfectly, but enough to notice the difference.

The Core Problem Nobody Talks About

Early on I just copied Skills into each agent’s directory. Seemed easier. Then I’d update a Skill in one place and forget the other three. Or I’d add a reference file in Claude Code’s Skill set and realize Codex was reading a stale version from last month. After a few weeks of this I genuinely did not know which Skill was the actual current one.

What I landed on: there can be many entry points, but only one source of truth.

One Source, Multiple Entry Points

I keep the Skills I actively use in ~/.agents/skills. Any agent that lets me add a custom load directory points there. For agents that need their own directory structure, I symlink back to the same root.

Update once, every connected agent picks it up. The catch: some agents have their own Skill loading logic and you need to check the docs before assuming they share. Took me a few hours to figure that one out.

Keep Skills Small and Focused

Don’t build a Skill that does research, writing, image generation, and review all at once. It becomes impossible to maintain. Change one part and you have no idea what else you broke.

A good Skill has one job with clear boundaries and something you can actually verify as done. If you can’t describe what a Skill produces in one sentence, it is probably doing too much. “Write an industry analysis” is not a Skill. “Pull facts from these three sources and cite them” is closer to something you can actually reuse.

Build Workflows, Not Mega Skills

Take researching a company. I break that into: collect data, clean the data, assess the business, compare against competitors, write the report. Each step is its own Skill. The workflow decides what runs first, what hands off to what, and where a human checks in.

When one step underperforms, I swap just that piece. When one step works really well, I can pull it into a sales prep workflow or a content research workflow. Same Skill, different assembly.

You do not need a complex workflow from day one. The simplest version is just: run Skill A, give output to Skill B, human checks the result before Skill C. Works fine.

Getting Started Without Overthinking It

If you want to get your Skills organized, do not start by designing a perfect directory structure. Pick one task you repeat every week. Break it into three to five steps. Define what goes in and what comes out of each step. Get one short workflow running, then extract the reliable parts into Skills.

A clean directory structure that nobody maintains is just an organized pile of prompts.

What I Actually Got From This

The real benefit was not saving disk space. It was stopping the low-grade anxiety of not knowing which version of a Skill an agent was running. I now spend less time confirming which file to update and more time actually building things.

Agents come and go. Entry points shift. What holds value is a set of Skills I can maintain, move between tools, and assemble as needed. That is the part worth investing in.