I vibe coded this at a hackathon a few weeks ago. The solution isn’t stable yet, but it’s similar to Claude Skills with the benefit of only needing an MCP client.

I open sourced it too.

It has a bit to go—like automatically monitoring reasoning traces and security—but I thought this would be easier to use than trying to maintain your own skills.md collection if the server handled more of it for you.

I’d be curious if anyone thinks I should keep working on it. I’d love someone to collaborate with on it!

I sometimes have to decide between running Claude Code headless with a prompt or let CC generate an application based on the prompt. (coining: Codification Threshold tradeoff)

The tradeoff is that the prompt is non-deterministic, but also more flexible. It can handle unknown situations. The generated code executes much faster and reliable for known situations.

If auto-learn would reduce the non-deterministic nature, i.e. chance of failing, that would speed up software development.

This is a very interesting concept. I have a few questions though.

Even if the agent has mcp tools avaible, you can't guarantee that it would necessarily use them right? I imagine that you'd have a lot of new tools that do nearly the same things.

Also how does this scale? If the agent has like 300 tools.

How does it determine whether the code it has for the skill, failed? (In order to determine that it needs to improve the code for that skill, I mean)

I don't understand, it'd be nice to see even a single example of the application