I kept bouncing between Gemini, Claude, and local models. Every single time, I'd rewrite the same tool definitions, the same system prompts, the same safety rules. It felt like teaching "what is a spreadsheet" to a new intern every week. I thought, "Why can't an agent just install a skill like I install a package?" What problem was I trying to solve? Fragmentation. Today, if you build a cool trading agent, nobody else can easily use that capability in their own agent running on a different model. Skills are trapped inside projects. I wanted a world where skills are portable—build once, run on any LLM, offline or online. How did my approach evolve? At first, I tried making a universal API wrapper. That failed because every model has different tool-calling quirks. So I pivoted to a skill bundle format: manifest.yaml (what the skill is), instructions.md (how the LLM should think), and skill.py (what actually runs). Then I built adapters that translate between the skill and each model's native language. The "aha" moment? Realizing the LLM doesn't need to understand the Python code—it just needs the instructions and the function schema. The framework handles execution. That's when it clicked. Result: One skill format. Any agent. Anywhere. Offline if you want. Repo: github.com/arpahls/skillware Site: skillware.site Would love to hear what skill you wish was just one pip install away. Contributors more than just welcome <3

Skillware

Plug-and-play skills for any AI agent.

Plug-and-play skills for any AI agent.

Forum Threads

What's the one skill you wish your AI agent already had?

Forum Threads

What's the one skill you wish your AI agent already had?