Imported from quantskills/skill-quant-factor-skill-factory (SKILL.md). Install upstream with npx skills add quantskills/skill-quant-factor-skill-factory. Copyright stays with the author (GPL-3.0-only).

{
  "version": 1,
  "task": {
    "placeholder": "请说明要生成或维护的 OHLCV 因子技能主题、现有索引与输出要求",
    "required": true
  },
  "fields": [
    {
      "key": "count",
      "label": "生成数量",
      "type": "number",
      "default": "10",
      "required": true
    },
    {
      "key": "start_id",
      "label": "起始因子 ID",
      "type": "number",
      "placeholder": "例如：1001"
    },
    {
      "key": "market",
      "label": "验证市场",
      "type": "select",
      "default": "both",
      "options": [
        { "value": "both", "label": "A 股与美股" },
        { "value": "cn", "label": "A 股" },
        { "value": "us", "label": "美股" }
      ]
    }
  ],
  "prompt_template": "{{#task}}任务与材料：\n{{task}}\n\n{{/task}}{{#attachments}}用户上传的材料（已放入工作区）：\n{{attachments}}\n\n{{/attachments}}使用量化因子技能工厂生成 {{count}} 个框架无关的 OHLCV 因子 Skill。{{#start_id}}从起始 ID {{start_id}} 起生成；{{/start_id}}未指定起始 ID 时按既有索引推断下一个可用编号。并在 {{market}} 真实缓存行情上验证；生成前检查既有索引避免重复，确保每个目录具备规定的技能说明、双语文档、因子与验证脚本、真实验证结果、公式参考和适配文件，核验通过数、数据供应商、市场覆盖、合并索引重复数及抽样验证结果，输出中文报告。"
}

Skill Quant Factor Skill Factory

Use this QuantSkills organization Skill when the user wants to create, extend, or maintain a library of framework-neutral quant factor Skills from OHLCV data.

The standard contract is:

• Factors are framework-neutral Python Skills for the QuantSkills organization.
• Users bring their own market data; generated factor code only requires open, high, low, close, volume, plus optional date, symbol, and market.
• Validation must use cached real OHLCV data when available. Do not describe synthetic validation as real validation.
• For China data, prefer AkShare A-share cache. For US data, prefer Yahoo Finance cache.
• Every generated factor folder must include SKILL.md, bilingual README.md, scripts/factor.py, scripts/validate.py, validation_real/result.json, validation_real/report.md, references/formula.md, and agents/openai.yaml.
• Before delivery, verify uniqueness across all previous factor indexes and confirm the generated package is complete.

Workflow

1. Inspect the current project state:

   Get-ChildItem
   Get-Content .\real_market_data\panels\panel_manifest.json

2. Confirm the real data panel contains market and vendor evidence:

   {
     "markets": {"cn": 98, "us": 50},
     "sources": ["akshare", "yahoo"]
   }

   Exact counts can vary by project, but the report must state the actual counts.

3. Generate the next batch with the reusable script.

   If the project already has tools/real_data_factor_pipeline.py, copy only scripts/generate_factor_skill_batch.py from this skill into the project's tools/ folder. If not, copy the validated pipeline scripts from a previous project first, then adapt data download paths.

   Example:

   $env:PYTHONUTF8='1'
   python .\tools\generate_factor_skill_batch.py `
     --count 200 `
     --start-id 1001 `
     --existing-index .\real_data_factor_skills_all_1000_index.json `
     --output-root .\real_data_factor_skills_extra_200_next `
     --combined-index .\real_data_factor_skills_all_1200_index.json `
     --report-name extra_200_next_factor_evaluation_report.md

4. Run validation and acceptance checks:

   python -m py_compile .\tools\generate_factor_skill_batch.py
   python .\tools\generate_factor_skill_batch.py --count <N> --start-id <ID> --existing-index <index.json> --output-root <folder> --combined-index <index.json> --report-name <report.md>

5. Audit the output:

   • Count generated factor directories equals the requested count.
   • validation_summary_real.json has the requested count.
   • All rows have status == "pass" unless the user explicitly accepts review rows.
   • All rows show the expected real-data market coverage.
   • Combined index has no duplicate slug.
   • Run scripts/validate.py for the first, middle, and last generated factor folders.

Output Summary

Always finish with:

• output folder
• factor ID range
• pass count
• real data panel counts and market vendors
• combined index count and duplicate count
• report path
• sampled scripts/validate.py result

Important Guardrails

• Do not claim Yahoo data is present unless real_market_data/raw/yahoo_us/*.parquet and the panel manifest prove it.
• Do not overwrite previous batches unless the user asks for regeneration.
• Use UTF-8 when reading or writing Chinese filenames and Markdown: set PYTHONUTF8=1 in PowerShell.