@qq-radio/skills

Install qq-radio skills into Codex, Claude, or Cursor on Windows.

Features

• Interactive installer: run qq-radio-skills
• Scriptable mode: --tool, --target-dir, --yes, --dry-run
• Doctor mode: qq-radio-skills doctor
• Idempotent overwrite for qq-radio-managed files only
• skills/ + manifest.json driven skill management

Repository layout

• Skill source content lives in skills/*.md
• Install metadata lives in root manifest.json
• The manifest source field uses repo-root-relative paths such as skills/vue-api-rule.md
• Promptfoo eval suites live in evals/skills/<skill-id>

Local development

pnpm install
pnpm build
pnpm link --global
qq-radio-skills doctor

Because the package ships a root bin wrapper, local pnpm link also works after prepare.

Promptfoo evals

This repo includes a Promptfoo-based compliance eval pipeline for skill markdown files.

• Each skill gets its own suite under evals/skills/<skill-id>
• Case data lives in evals/skills/<skill-id>/cases.json
• Prompt templates live in evals/skills/<skill-id>/prompts
• Custom rule assertions live in evals/assertions

Build the eval fixtures:

pnpm evals:build

Run the vue-api-rule suite with Promptfoo:

pnpm evals:vue-api-rule

Required environment variables:

• DASHSCOPE_API_KEY: DashScope API key used by the Promptfoo Alibaba provider
• DASHSCOPE_BASE_URL: compatible-mode endpoint, typically https://dashscope.aliyuncs.com/compatible-mode/v1
• DASHSCOPE_CHAT_MODEL: chat model id, for example qwen-plus
• DASHSCOPE_EMBEDDING_MODEL: not used by this skill compliance eval flow today

The eval flow is:

1. Read the skill markdown from skills/*.md
2. Expand suite cases into Promptfoo test cases
3. Ask the model to generate code while following the skill
4. Score each case with deterministic JavaScript assertions keyed by API-RULE-xxx

PowerShell example:

$env:DASHSCOPE_API_KEY = "your-api-key"
$env:DASHSCOPE_BASE_URL = "https://dashscope.aliyuncs.com/compatible-mode/v1"
$env:DASHSCOPE_CHAT_MODEL = "qwen-plus"
pnpm evals:vue-api-rule

When you add a new skill markdown file later, create a matching evals/skills/<skill-id>/suite.json and cases.json, then rerun pnpm evals:build.

Usage

qq-radio-skills
qq-radio-skills --tool codex
qq-radio-skills --tool claude --yes
qq-radio-skills --tool cursor --target-dir C:\Users\you\.cursor\rules --yes
qq-radio-skills --tool codex --dry-run
qq-radio-skills doctor

Default Windows target directories

• codex: %USERPROFILE%\.codex\skills
• claude: %USERPROFILE%\.claude\skills
• cursor: %USERPROFILE%\.cursor\rules

--target-dir always takes precedence over inferred defaults.

Installed output

• Codex: <targetDir>\<skillId>\SKILL.md
• Claude: <targetDir>\<skillId>\SKILL.md
• Cursor: <targetDir>\<targetFileName>.mdc

Generated files include a qq-radio header so repeat installs can overwrite only qq-radio-managed outputs.

doctor and --dry-run use the same manifest and skill sources as real installs; no extra flags are required for the new layout.

Publishing

The package is ready for pnpm link and can later be published to npm by updating the registry configuration in your release workflow.