跳到主要内容
    ↑↓ 选择↵ 打开esc 关闭
    antongulin

    Skill Creator

    opencode-skill-creator·v0.3.6·工具与命令

    OpenCode plugin for skill creation — custom tools for validation, evaluation, description optimization, benchmarking, and review serving.

    GitHub 星标

    188

    近 30 天 +19

    月装机量

    3,195

    近 7 天 485

    综合评分

    67.0

    生态多维模型

    最近提交

    1 天前

    2026-10-20

    快速安装与配置

    opencode.json

    写入当前项目的 opencode.json,只对这个仓库生效。

    opencode.json

    {
      "$schema": "https://opencode.ai/config.json",
      "plugin": ["opencode-skill-creator@0.3.6"]
    }

    OpenCode 启动时会通过内嵌运行时自动加载 npm 依赖并缓存至本地目录,无需手动在全局环境执行安装。

    OpenCode Skill Creator — Build better skills. Create. Evaluate. Improve.

    Create, evaluate, and improve OpenCode skills with a bundled workflow skill and TypeScript plugin.

    npm License: Apache 2.0 GitHub stars

    Website · Documentation · Examples · npm

    Quick start

    For OpenCode V2, run this from your project directory with Node.js and npm available:

    npx opencode-skill-creator install --project
    

    Restart OpenCode in that project, then ask:

    Use opencode-skill-creator to create a skill that reviews Docker Compose files.

    The project flag scopes plugin configuration to that project. At startup, the plugin still installs its bundled skill in the user-level OpenCode skills directory: $XDG_CONFIG_HOME/opencode/skills/opencode-skill-creator/ when set, otherwise ~/.config/opencode/skills/opencode-skill-creator/.

    For all projects, use install --global. OpenCode V1 ≥ 1.18.29 needs manual V1 setup. See installation for requirements, verification, and recovery.

    What it does

    • Draft a skill with SKILL.md frontmatter and supporting files.
    • Build should-trigger and should-not-trigger eval queries.
    • Measure description trigger accuracy and improve it with a train/test split.
    • Compare graded task outputs, timing, and tokens across runs.
    • Review paired outputs in an HTML viewer before installing your finished skill.

    This independent adaptation of Anthropic's skill-creator ports its Python automation to TypeScript for OpenCode. Trigger evals measure skill selection; use graded task outputs and human review to assess output quality. Model and provider behavior varies. The code is open source; model/provider usage can cost money.

    Install

    The installation guide covers project and global configuration, manual configuration without the CLI, compatibility, upgrades, and backup-first recovery. The installer targets V2's plugins key and preserves existing entries, including a legacy plugin array.

    OpenCode Configuration Setup
    V2 plugins Project quick start above or manual configuration
    V1 ≥ 1.18.29 plugin Supported V1 setup
    V1 before 1.18.29 Function default export Use the older 0.2.x line; the current bundle does not provide that entrypoint

    Plugin tools

    The plugin exposes 14 tools:

    Tool Purpose
    skill_validate Validate SKILL.md structure and frontmatter
    skill_parse Read name, description, and full content
    skill_eval Measure description trigger accuracy
    skill_improve_description Propose a description from eval failures
    skill_optimize_loop Evaluate and improve descriptions with a train/test split
    skill_aggregate_benchmark Summarize grading, timing, and token usage
    skill_generate_report Write an HTML description-optimization report
    skill_serve_review Start an HTTP review viewer
    skill_stop_review Stop review servers owned by this plugin instance
    skill_export_static_review Write a standalone HTML review
    skill_add_gold_standard Save a skill description example with its observed pass rate
    skill_list_gold_standards List all saved description examples
    skill_remove_gold_standard Remove a saved description example by ID
    skill_get_gold_advice Format saved description examples for prompt context

    See the tool reference for inputs, defaults, eval isolation, parser limits, and the review workflow.

    Usage

    After installation, ask OpenCode to use the workflow:

    • “Use opencode-skill-creator to build a skill for API documentation.”
    • “Use opencode-skill-creator to improve the description of my existing skill.”
    • “Create should-trigger and should-not-trigger queries, then evaluate this skill.”

    Examples

    Start with the Docker Compose helper for a complete small skill, eval-set JSON, and prompts for skill_eval, skill_optimize_loop, and benchmark review.

    Draft skills and eval workspaces in a system temp directory outside your source repository. Install the finished skill under .opencode/skills/<skill-name>/ for a project or the effective user-level OpenCode skills directory. The tool guide includes platform path examples.

    Architecture

    Component Responsibility
    Bundled skill Markdown workflow, grading/analyzing/comparison agents, schemas, and templates
    Plugin Validation, trigger evals, description optimization, benchmark aggregation, and review tools

    OpenCode loads the compiled dist/skill-creator.js bundle through its V1 or V2 entrypoint. Plugin startup installs or updates the bundled skill at user level, even with project configuration. Updates attempt to retain a customized SKILL.md but refresh bundled support files. Read upgrade and recovery guidance before resetting or restoring files.

    Source map

    • opencode-skill-creator/: source skill, agents/, references/schemas.md, and templates/eval-review.html.
    • plugin/skill/: bundled copy of that skill; keep it aligned with the source.
    • plugin/skill-creator.ts and plugin/runtime-entry.ts: tool registration and runtime entrypoints.
    • plugin/lib/: parsing, validation, evals, optimization, aggregation, installation, and review servers.
    • plugin/templates/viewer.html: task-output review viewer; plugin/dist/: build output shipped on npm.
    • examples/: runnable workflow examples; docs/: installation, tool reference, and publishing assessment.

    Differences from the Anthropic original

    The original workflow uses Claude CLI calls and Python scripts. This version invokes opencode run for evals and description improvement, exposes tools such as skill_optimize_loop and skill_aggregate_benchmark, and serves review through skill_serve_review. Distribution uses an npm plugin plus a skill directory rather than a .skill archive. It targets OpenCode; Anthropic provenance does not imply support for other clients.

    Contributing

    Read CONTRIBUTING.md for development checks and SECURITY.md for private vulnerability reports. Report ordinary bugs in GitHub issues.

    Built by Anton Gulin. Sponsor the project.

    License

    Apache License 2.0. Based on anthropics/skills by Anthropic.

    同类生态推荐