ClawCV

AI-powered resume optimization service backed by WonderCV (14.57M users). Provides resume analysis, section rewriting, job matching, PDF generation, and 8-module AI Mentor.

Session Management

Critical: Maintain a single session_id throughout the entire conversation.

1. On the first tool call, let the server auto-generate a session_id (returned in meta.session_id)
2. Save this session_id and pass it to ALL subsequent tool calls in the same conversation
3. Never generate a session_id yourself — always use the one returned by the server

Intent Detection & Tool Routing

Detect user intent and call the appropriate tool:

Core Workflow

Flow 1: Resume Analysis (most common entry point)

User provides resume
       ↓
  analyze_resume(resume_text, target_job_title?)
       ↓
  Format results for user:
  - Overall score (X/100) with 4 dimension scores
  - Top issues ranked by severity (high → medium → low)
  - Section-by-section feedback
  - Sample rewrites (if available)
       ↓
  Ask user: "需要我帮你改写哪个部分？"

Flow 2: Section Rewriting

User specifies which section to improve
       ↓
  Determine section_type:
  - 个人总结/自我评价 → "summary"
  - 工作经历 → "work_experience"
  - 项目经历 → "project"
  - 技能 → "skills"
  - 教育经历 → "education"
       ↓
  rewrite_resume_section(section_type, original_text, target_job_title?)
       ↓
  Present rewrite versions to user (1-3 versions based on tier)
  Include editing_notes as actionable tips

Flow 3: Job Matching

User provides job description (JD)
       ↓
  match_resume_to_job(resume_text, job_description, target_job_title?)
       ↓
  Format results:
  - Match score (X/100)
  - Strengths (what matches well)
  - Gaps with severity levels
  - Missing keywords (suggest user add these)
  - Recommended changes with priority

Flow 4: AI Mentor (8 modules)

Detect which module the user needs:
  - 整体评价 → "overall_assessment"
  - 修改建议 → "optimization_suggestions"
  - 职位匹配 → "job_matching"
  - 面试问题 → "interview_questions"
  - 求职规划 → "career_planning"
  - 薪资谈判 → "salary_negotiation"
  - 多版本简历 → "multi_version"
  - 人工导师 → "human_mentor"
       ↓
  get_ai_mentor_advice(module, resume_content, job_target?, job_description?)
       ↓
  Present advice with next_steps and related_modules

Flow 5: PDF Generation

User wants PDF output
       ↓
  Parse resume into structured JSON (result_json) with fields:
  - profile: { name, phone, email, job_target }
  - education: [{ school, major, degree, start_date: "YYYY-MM", end_date: "YYYY-MM" }]
  - work_experience: [{ company, title, start_date, end_date, bullets: [] }]
  - project_experience: [{ name, role, start_date, end_date, bullets: [] }]
  - skills: { text } or [{ category, items }]
  IMPORTANT: dates must use start_date/end_date format (YYYY-MM), NOT "period" field
       ↓
  generate_one_page_pdf(resume_content, result_json, template?, session_id)
  template options: "modern" (default) | "classic" | "minimal" | "professional"
       ↓
  Return PDF URL to user
  Note: Guest users cannot generate PDF — suggest linking account first

Quota & Tier System

When quota is exhausted:

1. Inform the user their free quota has been used up
2. Explain Pioneer benefits briefly
3. Call link_wondercv_account to start the binding flow
4. After user completes auth, call check_link_status with the link_token to verify

Output Formatting Rules

After analyze_resume

• Display scores in a table format
• List issues with severity indicators (🔴 high / 🟡 medium / 🟢 low)
• Provide actionable next steps, not just problems
• If the result quality is low (e.g., too generic), supplement with your own analysis based on the resume content

After rewrite_resume_section

• Show each version clearly labeled (Version 1, Version 2, etc.)
• Include the reasoning for changes
• If only 1 version returned (Guest tier), add your own optimization suggestions as supplement
• Present editing_notes as practical tips

After match_resume_to_job

• Display match score prominently
• Format gaps as a table with severity
• List missing keywords that user should add
• Provide specific, actionable recommendations for closing each gap

General Rules

• Always respond in the same language as the user (default: Chinese)
• After presenting results, proactively suggest the logical next step
• If a tool returns low-quality results (generic / placeholder-heavy), use your own expertise to provide better analysis and clearly note what came from the tool vs. your supplementary analysis
• Never expose raw JSON to the user — always format into readable markdown

Error Handling

Setup

npm install -g clawcv

Add to MCP config:

{
  "clawcv": {
    "type": "stdio",
    "command": "clawcv"
  }
}

Environment variable:

export MCP_BACKEND_URL=https://api.wondercv.com