yintian-skill

name

yintian-skill

description

本地身份信息保险柜与自动填表协调 Skill。当用户想用身份证、银行卡、户口本、简历、证件照片、扫描件、PDF/Word/Excel/文本等本地私密材料，自动填写网页表单、Word/PDF/Excel 文档、App 表单、注册页或入职登记表时，使用本 skill。它在本地提取所需身份字段，图片用 RapidOCR + OpenVINO 本地 OCR，敏感身份数据不进入远程大模型/在线 AI 上下文，保留来源证据，把实际填写动作交给浏览器/文档/PDF/表格等现成工具，并要求最终给出已填、缺失、需确认、冲突的验收报告；它本身不实现网页或文档填写模块。 Local private identity vault and autofill coordination skill. Use this whenever the user wants to automatically fill a website, Word/PDF/Excel document, app form, registration page, onboarding form, or other target with identity information stored in local private materials such as ID cards, bank cards, household documents, resumes, scans, screenshots, PDFs, Word/Excel/text/JSON/CSV files, or document photos. This skill extracts needed identity fields locally, uses RapidOCR + OpenVINO local inference for image OCR when needed, keeps sensitive identity data out of remote LLM/online AI contexts, preserves source evidence, coordinates handoff to the appropriate browser/document/PDF/spreadsheet/tooling skill for the actual filling, and requires a final acceptance report listing filled fields, missing fields the user must provide, ambiguous/conflicting values, and any fields needing confirmation. It should not implement website/document/form filling modules itself.

Yintian Identity Vault And Autofill Coordinator

Use this skill as the user's local identity vault plus the process harness around identity-based autofill.

The skill answers three questions:

1. What identity information can be extracted from the user's local private files?
2. Which extracted fields should be handed to another tool/skill for the target being filled?
3. After the downstream fill attempt, what was filled, what is missing, and what needs user confirmation?

It does not implement the downstream writer. Websites, Word files, PDFs, spreadsheets, and app forms should be handled by the appropriate browser, document, PDF, spreadsheet, or other available tool.

Scope

This skill owns:

• Reading user-provided local private files.
• Enforcing the optional vault directory boundary with YINTIAN_VAULT_DIR.
• OCR for document images, ID-card photos, scans, and screenshots through RapidOCR + OpenVINO local inference.
• Extracting identity-field candidates such as name, ID number, birth date, gender, address, phone, bank card, and bank name.
• Mapping extracted identity candidates to the target's requested fields.
• Coordinating handoff to the right downstream capability.
• Requiring a final acceptance report with filled, missing, ambiguous, and confirmation-needed fields.

This skill does not own:

• Writing into websites, Word/PDF/Excel files, or app forms.
• Clicking submit.
• Reimplementing browser automation, document editing, spreadsheet editing, or PDF editing.
• Sending private source text, OCR text, ID numbers, bank cards, addresses, or full identity packets to remote LLMs, online AI services, or network paths unrelated to the user-selected target.

Required Workflow

1. Understand target requirements. Inspect or infer what identity fields the target needs. If the target is a website/document/PDF/spreadsheet, use the appropriate external tool/skill to inspect it.
2. Extract identity packet. Use scripts/source_extract.py and scripts/identity_extract.py to read local private materials and build a field packet with evidence.
3. Map fields. Create a target mapping: target field -> selected identity value -> source evidence. Keep unresolved fields explicit.
4. Handoff. Ask the agent to use the appropriate downstream tool/skill to perform the actual fill. Do not use this skill to implement the fill.
5. Acceptance check. After filling, verify by reading back the target if possible. If readback is not possible, require a fill ledger from the downstream action.
6. Report to user. Always report filled fields, missing fields, ambiguous/conflicting values, low-confidence values, and fields requiring user supplement or confirmation.

For detailed guidance, read references/privacy-extraction-workflow.md.

Output Contracts

Private Data Packet

{
  "fields": {
    "name": [
      {
        "value": "张三",
        "source": "/vault/id-card.txt",
        "evidence": "姓名 张三",
        "confidence": "label"
      }
    ]
  },
  "derived": {
    "birth_date": {
      "value": "1985-03-07",
      "source": "/vault/id-card.txt",
      "evidence": "derived from id_number"
    }
  },
  "needs_review": []
}

Target Mapping

{
  "target": "/path/or/url",
  "mapping": [
    {
      "target_field": "出生日期",
      "identity_field": "birth_date",
      "value": "1985-03-07",
      "source": "/vault/id-card.txt",
      "status": "ready"
    }
  ],
  "missing": [
    {
      "target_field": "紧急联系人",
      "reason": "not found in local private materials"
    }
  ],
  "needs_confirmation": []
}

Final Acceptance Report

{
  "filled": ["姓名", "身份证号", "出生日期"],
  "missing_user_input": ["紧急联系人"],
  "needs_confirmation": ["OCR 低置信度的开户地址"],
  "conflicts": [],
  "downstream_tool": "browser/document/pdf/spreadsheet tool used by agent",
  "readback_verified": true
}

The final report is part of this skill's process responsibility. Missing identity information should be shown to the user at the end instead of being invented.

Scripts

• scripts/source_extract.py: Read local .txt, .md, .json, .csv, .tsv, .pdf, .docx, .xlsx, and common image files. Images go through local OCR.
• scripts/identity_extract.py: Convert extracted text into identity-field candidates and derived fields.
• scripts/ocr_engine.py: Run local OCR through RapidOCR/OpenVINO. Device is controlled by YINTIAN_OCR_DEVICE=CPU|GPU|NPU|AUTO.
• scripts/privacy.py: Enforce the optional vault path boundary and detect sensitive strings.
• scripts/mcp_server.py: Expose vault extraction and OCR as MCP tools.
• scripts/demo.py: Run an offline text demo without OCR dependencies.

Downstream Handoff Rules

• Website target: use browser/browser-control tools after identity extraction. Ask before submitting.
• Word target: use a Word/document skill or tool after identity extraction.
• PDF target: use a PDF skill or tool after identity extraction.
• Spreadsheet target: use spreadsheet/xlsx tooling after identity extraction.
• Unknown target: choose a downstream capability based on the target, not by adding new fill logic to this skill.

Keep this skill focused: it provides private identity data, coordinates the autofill workflow, and checks the result; other tools perform target-specific writing.

Quick Commands

Extract raw local text:

python yintian-skill/scripts/source_extract.py /path/to/private-file.pdf

Build a private identity packet:

python yintian-skill/scripts/identity_extract.py /path/to/id-card.txt /path/to/bank-card.jpg

Run the MCP server:

python yintian-skill/scripts/mcp_server.py