Hey everyone! 👋 I'm Anton, the creator of opencode-skill-creator.

I built this because I was frustrated with how hit-or-miss AI agent skills are. You write a skill, test it manually, maybe tweak the description a few times, and hope it works. There was no systematic way to evaluate whether a skill triggers correctly or to measure improvements across iterations.

When I saw Anthropic's skill-creator for Claude Code, I loved the methodology — eval-driven development for AI skills — but it only worked for Claude Code and required Python. So I ported it to TypeScript and packaged it as an OpenCode plugin that anyone can install with one command.

The key insight: skills are software, and software should be tested. The description optimization loop alone was a game-changer for my own skills — it takes a skill from "maybe it'll trigger" to "quantitatively proven to trigger on the right prompts."

How it works:

You describe what skill you want (or use an existing one)
The tool generates test cases automatically
It runs evals — with and without the skill — to measure triggering accuracy
An LLM-powered optimization loop iteratively improves the skill's description
A visual review viewer lets you evaluate quality as a human
You benchmark results with variance analysis across iterations

Technical details:

TypeScript plugin with zero Python dependencies
Registers custom tools in OpenCode (skill_validate, skill_eval, skill_optimize_loop, etc.)
Based on Anthropic's official skill-creator architecture
Apache 2.0 license

Happy to answer any questions! Also looking for contributors and feedback from OpenCode users.

GitHub: https://github.com/antongulin/opencode-skill-creator

npm: https://www.npmjs.com/package/opencode-skill-creator