self-building skills

Skills that don't break
when you share them.

You got Claude doing something just right. You share it, and for everyone else it falls flat, because it quietly leaned on your setup. skillc is a skill that makes skills: tell it what you want, show it a couple of examples you like, and it hands you one file you can send to anyone. Their Claude sets it up for itself.

Get skillc install once, then just ask
1
Turn on Code Execution. Settings, then Capabilities. A one-time toggle, any plan.
2
Download skillc:
3
Upload it. Settings, then Capabilities, then Skills, then drop the zip in.

Paste into Claude Code.

OpenCode reads ~/.claude/skills/ automatically. Get skillc.zip first.

The fallback if you would rather not install. Same builder, runs in a chat.

Then just ask: skillc, make me a skill that writes my emails in my voice. It builds you a skill; send it to anyone and they install it the same way.

what do you do with it

Watch the whole loop.

You ask for a skill and show it what good looks like. It tries one, you approve or fix the test, and it hands you a file. Send that file to anyone; it sets itself up for whoever opens it.

assistantskillc

how it works

The file rebuilds itself.

A skill made by skillc does not carry finished instructions. It carries the examples you approved, plus a recipe for turning them into instructions. The first time someone uses it, before it answers:

the file

One file, sent

Your approved examples plus the rebuild recipe. Nothing else to ship.

on open

Rebuilds

Their Claude writes itself instructions that reproduce your examples, on their setup.

checks

Grades itself

On a few approved examples that were held back from the rebuild.

result

Runs, or says so

Reports a score and runs, or says plainly it could not build. Never quietly wrong.

measure it yourself

Every skill reports its own transfer score.

Every time a skill sets itself up somewhere new, it grades itself on those held-back examples and prints a score from 0 to 1. The same file on a different model gets you a fresh number, no setup. Here is what that produced for the example skills:

0.93
Voice skill, run in a session told to write in the opposite style. It held its style. Transfer score 0.93 against the original.
18 of 18
Build-then-use across every Opus, Sonnet, and Haiku maker-receiver pairing, two skills. All eighteen rebuilt in spec.
0 of 6
Six poisoned instructions tried to hide a step in the built skill. None reached the output.

Every run is logged in the examples, method and inputs and result. It is all in the open: rebuild any of them on your own model, then send the score back. Open an issue with what you saw, or a pull request that adds examples or tightens a skill's checks. Those go into the file and ship to everyone.

for developers

It is a marketplace, too.

The Claude Code install above uses skillc's plugin marketplace, so you can version skills and ship them as files. The worked example skills are a separate plugin, skillc-examples (voice, explain-plainly, commit-message): install with /plugin install skillc-examples@skillc. One honest note: on claude.ai, uploaded skills only share through the platform on Team or Enterprise plans, so sending the file (or the paste path) is what works for everyone. The format, the builder, and the rebuild recipe are all in the repo.