Skip to content

pentaxis93/skill-creator

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

16 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

skill-creator (platform-agnostic fork)

Create new skills, improve existing skills, and measure skill performance.

What this is

A fork of anthropics/claude-plugins-official (plugins/skill-creator/) that produces platform-agnostic skills — usable by any agent that reads SKILL.md files, not just Claude Code.

The tool itself still runs on Claude Code (that's fine — it's the authoring environment). Only the generated output (SKILL.md files, descriptions, reference content) has been neutralized to remove platform-specific assumptions.

Installation

Prerequisites

  • Python 3.8+
  • claude CLI — required by run_eval.py and run_loop.py

Note: Only tested on Debian. Other Linux distributions, macOS, and Windows are untested.

Python dependencies

python3 -m venv .venv
source .venv/bin/activate
pip install anthropic pyyaml
Package Required by Purpose
anthropic improve_description.py, run_loop.py API calls for description optimization
pyyaml quick_validate.py, package_skill.py SKILL.md frontmatter parsing

All other scripts use Python standard library only.

What changed

Two files modified, everything else untouched:

  • skills/skill-creator/SKILL.md — All output-shaping and contextual references neutralized, plus a new "Platform-Agnostic Output" section. Key replacements: "Claude" → "the agent" / "AI" / "an LLM", "MCPs" → "tools", "Claude's available_skills list" → "the agent's skill list", "Claude.ai-specific instructions" → "Limited-environment instructions", "Cowork" → generic multi-agent terminology. The claude CLI binary name is retained where it refers to a shell command.

  • skills/skill-creator/scripts/improve_description.py — Prompt template neutralized. Fork note at module level to reduce merge-conflict surface with upstream.

Files not modified (confirmed no output-shaping Claude references): run_eval.py, run_loop.py, aggregate_benchmark.py, generate_report.py, package_skill.py, quick_validate.py, utils.py, grader.md, comparator.md, analyzer.md, schemas.md, viewer.html, generate_review.py, eval_review.html.

Upstream source

Repository: https://github.com/anthropics/claude-plugins-official
Subtree:    plugins/skill-creator/
License:    Apache 2.0

Syncing upstream changes

Requires git-filter-repo. To pull in upstream updates:

UPSTREAM_DIR="$(mktemp -d)"

# Fresh extraction of upstream subtree
git clone https://github.com/anthropics/claude-plugins-official.git "$UPSTREAM_DIR"
git -C "$UPSTREAM_DIR" filter-repo --subdirectory-filter plugins/skill-creator/

# Merge into this repo (run from your skill-creator checkout)
git remote add upstream-update "$UPSTREAM_DIR"
git fetch upstream-update && git merge upstream-update/main
git remote remove upstream-update
rm -rf "$UPSTREAM_DIR"

Resolve any conflicts in SKILL.md or improve_description.py by keeping the platform-agnostic wording.

License

Apache 2.0 — see LICENSE.

About

Anthropic's skill-creator, modified to produce vendor-agnostic skills. Runs on Claude Code, outputs skills any agent can use.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors