ds160-autofill（ds160 自动填充）

DS-160 Auto-fill Skill

This skill automates the filling of US nonimmigrant visa DS-160 forms by combining:

• CDP (Chrome DevTools Protocol) for fast element location
• CSV data source for user information
• LLM assistance for complex cases (captcha, missing elements)
• Session persistence for resume capability
• Chinese to English translation for user input

Quick Start

Starting a New Application

1. Provide CSV data: User must provide a CSV file with their personal information (use ds160-user-info.csv as template)

1. Open DS-160 website: Use browser tool to open https://ceac.state.gov/genniv/ with profile openclaw

1. Initialize session: Load and execute scripts/ds160-filler.js to parse mappings and CSV

1. Start filling: Begin with the home page, handle captcha with LLM, then proceed page by page

Resuming an Application

1. Load existing session: Check ds160/ds160-session.json for saved Application ID

1. Open application: Use the saved Application ID to resume on DS-160 website

1. Continue filling: Load CSV from saved location and continue from last page

Core Workflow

Step 1: Prepare Data

Always check if session exists first:

// Read ds160/ds160-session.json
// If exists → Resume mode
// If not → New application mode

New application mode:

• Ask user for CSV file path or content
• Save CSV to ds160/ds160-user-info.csv using write tool
• Initialize session data in ds160/ds160-session.json:

  {
    "applicationId": null,
    "securityQuestion": null,
    "securityAnswer": null,
    "currentPageIndex": 0,
    "completedPages": [],
    "startDate": "2026-02-06T21:00:00Z"
  }

Resume mode:

• Load CSV from ds160/ds160-user-info.csv
• Load session from ds160/ds160-session.json
• Report current progress to user

Step 2: Initialize Browser

// Open browser with openclaw profile
browser_start: { profile: "openclaw", targetUrl: "https://ceac.state.gov/genniv/" }

// For resume mode, add Application ID to URL
// Example: https://ceac.state.gov/GenNIV/Common/ConfirmApplicationID.aspx

Step 3: Fill Current Page

For each page in sequence:

1. Snapshot page to understand current state:

   browser_snapshot: { refs: "role", profile: "openclaw" }

1. Load mappings from references/ds160-elements.yaml

1. Execute fill logic using browser act with evaluate:

   // Load ds160-filler.js
   // Call fillPage(page, currentUrl, userData, yamlData)

1. Handle results:

- Success: Continue to next page - Needs LLM assistance: Call LLM and retry - Needs user input: Save progress and pause

Step 4: Handle Special Cases

Captcha:

• Take screenshot of captcha area
• Use image tool to analyze captcha
• Fill captcha code via browser evaluate

Missing element:

• Call LLM with page snapshot
• Ask LLM to locate element and provide alternative selector
• Retry fill operation

Translation needed:

• Call LLM with field context and Chinese value
• Ask LLM to translate to appropriate English value
• Update CSV and retry fill operation

Missing user data:

• Identify which field is missing
• Report to user with field name and description
• Important: Save current page using browser evaluate before pausing
• Wait for user to provide data

Step 5: Save Progress & Continue

After completing a page:

1. Update session data:

   {
     "applicationId": "AA00FBLCQP",
     "securityQuestion": "What is the given name of your mother's mother?",
     "securityAnswer": "LiMei",
     "currentPageIndex": 5,
     "completedPages": ["home", "security_question", "personal_1", "personal_2"],
     "startDate": "2026-02-06T21:00:00Z"
   }

1. Save to file: write(ds160/ds160-session.json, JSON.stringify(sessionData))

1. Report progress to user:

- Current page completed - Next page name - Application ID - Security question/answer (first time only) - Overall progress (X/Y pages)

1. Click "Continue" button to proceed to next page

Data Structures

CSV Format

The CSV must have these columns:

1. 页面 - Page name
2. 字段名称 - Field identifier (matches YAML name field)
3. 英文说明 - English description
4. 中文说明 - Chinese description
5. 必填 - Required? (是/否)
6. 示例值 - Example value
7. 用户填写 - User-provided value (this is what gets filled)

Important: Users can provide values in Chinese. The script includes a built-in translation dictionary for common fields (gender, marital status, countries, etc.). For fields not in the dictionary, LLM will be called for translation automatically.

YAML Structure

Each page in ds160-elements.yaml contains:

• page_id: Unique page identifier
• page_name: Human-readable page name
• url: Page URL pattern
• elements: Array of form elements with:

- id: Element ID - name: Field identifier (matches CSV) - type: Element type (text, select, radio, checkbox, button) - label: English label - label_cn: Chinese label - required: Is required? - options: Available options (for select) - group: Radio group name

Error Handling

Element Not Found

Symptoms: fillPage returns needsLLM: true

Action:

1. Take page snapshot
2. Call LLM with:

- Current URL - Element information (id, name, type) - Page HTML snippet - Ask LLM to analyze and provide alternative selector

1. Retry with LLM-suggested selector

Translation Needed

Symptoms: fillPage returns needsTranslation: true

Action:

1. Call LLM with translation context:

- Field name and description (EN/CN) - Field type and available options - Original Chinese value - Ask LLM to translate to appropriate English value for the field

1. Update CSV with translated value
2. Retry fill operation with translated value

Example LLM prompt for translation:

I need to translate this Chinese value for a DS-160 form field:

Field: ${elementName}
Type: ${elementType}
English label: ${label}
Chinese label: ${label_cn}
Chinese value: "${originalValue}"

Available options (if select): ${options}

Please provide the correct English value for this field. Consider:
- The field type and context
- Available options (if applicable)
- Standard DS-160 terminology

Return only the English value, no explanation.

Data Missing

Symptoms: fillPage returns needsUserInput: true

Action:

1. CRITICAL: Save current page before pausing:

   // Find and click "Save" button
   browser_act: {
     request: { kind: "click", ref: "save_button" }
   }

1. Report missing field to user:

- Field name (English and Chinese) - Description - Whether it's required

1. Update CSV with new data using write tool
2. Resume filling

Captcha

Symptoms: Captcha image detected on page

Action:

1. Take screenshot of captcha area
2. Use image tool:

   image: {
     image: "/path/to/captcha.png",
     prompt: "What is the code shown in this captcha? Return only the code characters."
   }

1. Fill captcha field via browser evaluate
2. Submit form

Page Sequence

The pages are filled in this order:

1. home - Start application, select location, handle captcha
2. security_question - Confirm Application ID, set security question
3. personal_1 - Personal information (name, gender, DOB)
4. personal_2 - Nationality, other nationalities, IDs
5. travel - Travel purpose, dates, who is paying
6. travel_companions - People traveling with you
7. previous_us_travel - Previous US visits
8. address_phone - Current address and phone
9. passport - Passport information
10. us_contact - US point of contact
11. family_relatives - Family information
12. work_education - Work and education history
13. security_background - Security questions

Progress Reporting

Report to user after each page:

✓ Completed: Personal Information 1
📄 Next: Personal Information 2
🆔 Application ID: AA00FBLCQP
🔐 Security Question: What is the given name of your mother's mother?
🔑 Answer: LiMei
📊 Progress: 3/13 pages completed

Include in every report:

• Application ID (once obtained)
• Security question and answer (once generated)
• Current progress (X/Y pages)
• Next page name

LLM Integration

When to Call LLM

1. Captcha detection: Always use LLM to read captcha
2. Element not found: When multiple locator strategies fail
3. Translation needed: When user provides Chinese value not in built-in dictionary
4. Complex elements: For elements that require understanding context
5. Validation errors: When form submission fails with unclear reason

LLM Prompts

For element location:

I need to find this form element on a DS-160 page:
- Element ID: {elementId}
- Field name: {fieldName}
- Element type: {elementType}
- Current URL: {url}

Please analyze the page and provide:
1. The correct way to locate this element (CSS selector or JavaScript)
2. Any visible labels or text associated with it

For captcha:

Please read the captcha code from this image. Return only the code characters, no explanation.

Translation Dictionary

The script includes a built-in translation dictionary for common Chinese values. This covers:

Gender:

• 男/男性 → MALE
• 女/女性 → FEMALE

Marital Status:

• 已婚 → MARRIED
• 未婚 → SINGLE
• 离异 → DIVORCED
• 丧偶 → WIDOWED
• 合法分居 → LEGALLY SEPARATED
• 事实婚姻 → COMMON LAW MARRIAGE
• 民事结合 → CIVIL UNION/DOMESTIC PARTNERSHIP

Yes/No:

• 是/有/是，有 → Yes
• 否/没有/否，没有 → No

Countries:

• 中国 → CHINA
• 美国 → USA
• 英国 → UNITED KINGDOM
• 日本 → JAPAN
• 韩国 → KOREA, SOUTH

Visa Types:

• 旅游/商务/商务/旅游 → B1/B2
• 学生 → F1
• 访问学者 → J1
• 工作 → H1B

Payment:

• 自己/自费 → SELF
• 公司 → OTHER COMPANY
• 其他人 → OTHER PERSON

Relationship (US Point of Contact):

• 朋友 → Friend
• 亲戚 → Relative
• 同事 → Colleague
• 配偶 → Spouse
• 同学 → Classmate
• 其他 → Other

Other Values:

• 不适用/不，不适用 → Does Not Apply

For values not in this dictionary, LLM will be automatically called for translation.

Session Persistence

Session file location: ds160/ds160-session.json

Session data structure:

{
  "applicationId": "AA00FBLCQP",
  "securityQuestion": "What is the given name of your mother's mother?",
  "securityAnswer": "LiMei",
  "currentPageIndex": 5,
  "completedPages": ["home", "security_question", "personal_1", "personal_2", "travel", "travel_companions"],
  "startDate": "2026-02-06T21:00:00Z",
  "lastUpdated": "2026-02-06T21:30:00Z"
}

When to save session:

• After completing each page
• Before pausing for user input
• After receiving Application ID
• After setting security question/answer

Browser Automation Notes

Always use openclaw profile to maintain session across pages.

Use browser act with evaluate for form filling:

• More reliable than individual click/type actions
• Can execute complex JavaScript
• Better error handling

Use browser snapshot before each page:

• Check page state
• Identify dynamic elements
• Verify page loaded correctly

Testing Checklist

Before using this skill with real data:

• [ ] Test element location on each page type
• [ ] Verify CSV parsing with sample data
• [ ] Test captcha handling with LLM
• [ ] Verify session save/load functionality
• [ ] Test resume from middle of application
• [ ] Verify "Save" button clicking works before pauses
• [ ] Test with various element types (text, select, radio, checkbox)
• [ ] Verify progress reporting format
• [ ] Test Chinese to English translation

Example Usage

User says: "Help me fill DS-160 form with my data in this.csv"

Agent workflow:

1. Load SKILL.md (this file)
2. Check for existing session
3. Save CSV to ds160/ds160-user-info.csv
4. Initialize session
5. Open browser to DS-160
6. Fill page by page
7. Report progress after each page
8. Handle errors with LLM or user input
9. Save progress and continue

User says: "Continue filling my DS-160 form"

Agent workflow:

1. Load SKILL.md
2. Load existing session
3. Open browser to DS-160 with Application ID
4. Load CSV from saved location
5. Continue from last page
6. Report progress
7. Complete remaining pages

Files Reference

• scripts/ds160-filler.js - Core automation logic (run via browser evaluate)
• references/ds160-elements.yaml - Element mappings (read-only)
• references/ds160-user-info.csv - User data template (copy to workspace)
• ds160/ds160-session.json - Session persistence (auto-created)
• ds160/ds160-user-info.csv - Active user data (auto-created)

Important Notes

• NEVER submit the final form - this is for automation/testing only
• Always save before pausing - prevents data loss
• Report progress frequently - at least every page
• Keep Application ID secure - it's sensitive information
• Use fake data for testing - don't use real personal info
• Chinese input is supported - use built-in dictionary or LLM for translation