Skilldocs

Figma for markdown

Productivity

Developer Tools

Artificial Intelligence

Visit WebsiteSee on Product Hunt

This is awesome! I have the wrong mental model for the MCP server install, though, and could use some help. I clicked on the "Claude Code can work in here" and am following the instructions. To install the MCP server using this command, do I need to check out a local copy of skilldocs source code? (Very sorry if this is obvious in docs that I should have seen.)

```
claude mcp add skilldocs -e SKILLDOCS_TOKEN=<paste-token> -- node <path-to-skilldocs-repo>/mcp/index.js
```

The versioning thread above is the right conversation, so let me bring a different problem.

I author skills as markdown, keep them in a private repo, and push them to a second machine (a Windows box) where a different agent loads and runs them. What I have never been able to see is whether an edit made a skill better. I can tell instantly whether the prose improved. I cannot tell whether the agent followed it more reliably afterward, and those two come apart more often than I expected. I have made a skill clearer to read and worse to obey.

So collaborative editing closes the human-to-agent loop for edits, and it looks great. But the loop I need closed runs the other way: this version of the skill, run N times, produced these outcomes.

Is there any notion in Skilldocs of a version carrying results, even something as crude as letting me mark "that run went well" against whichever version was live? Gal's point about telling an agent to use "the one from Tuesday's meeting" gets much stronger if Tuesday's version also carries evidence that it actually worked.

on the source control question - since everyone's already live in the same doc like Figma, I'm not sure you actually want git branches, branching is for when people edit in isolation and need to reconcile later, which is the opposite of what you've built. what I'd actually want is lighter: the ability to name/tag a version before I hand the diff to my coding agent, so if the agent runs with a bad version of the skill I can point back to "the one from tuesday's meeting" instead of reconstructing it from memory. real branching only starts to matter once two agents are independently proposing edits to the same skill without a live human session tying them together, which sounds like a later problem, not a v1 one.

Skilldocs solves a surprisingly awkward collaboration gap! How are you thinking about source control once multiple people and agents are editing the same skill? Will comments and diffs eventually map cleanly back to Git branches/PRs?