Resume Tailor

Analyze a job description and tailor the source resume to match, with interactive checkpoints for content decisions and full reverse audit.

Trigger Phrases

• "帮我针对这个 JD 调简历" / "tailor resume to this JD"
• "优化简历给这个岗位" / "optimize resume for this role"
• "这份简历匹配这个岗位吗" / "is my resume a good fit"
• Any JD text or URL provided with resume adjustment intent

Resume Input

On first run, ask the user to provide their resume file path (absolute path to .docx, .pdf, or .txt). Acceptable formats:

• .docx — preferred, preserves formatting
• .pdf — readable but harder to preserve layout
• .txt / .md — plain text

Store the path in memory for the session. If the user provides a new path, update accordingly.

All tailored versions derive from this source file. Never modify the original.

If no path is provided and no resume file is found in the workspace, guide the user to place one.

---

Five Core Principles / 五大核心原则

These are the non-negotiable foundations of this skill. Every phase and checkpoint must uphold them.

1. Hybrid Analysis / 混合分析

Script (jd_parser.py) extracts structured features (years, degrees, certifications). LLM extracts semantic meaning (capability clusters, business logic). Both cross-validate — script catches what LLM misses, LLM interprets what script can't.

2. Fact Conservation / 事实守恒

• Strict reverse chronological order: Work/internship entries are NEVER reordered by relevance. Only inclusion or exclusion is allowed.
• Structure recovery confirmation: When resume styling is lost and LLM reconstructs section boundaries, the user MUST confirm before proceeding.

3. Semantic Equivalence / 语义等价

• Cross-credential mapping: IELTS 7.0 ≈ CET-6 550+ — retain the original credential and add an equivalence note.
• Cultural tone slider: Adjust assertiveness level based on target region (North America: assertive; East Asia: collaborative; etc.). Applied in CP4 as suggestions, never auto-applied.

4. Closed-loop Quantification / 闭环量化

Anti-Filler Rule (see interaction_checkpoints.md): If a bullet lacks a quantified outcome, the LLM must progressively probe (up to 2 rounds) for measurable evidence. If the user still cannot provide data, keep the original wording or remove the outcome clause. A vague outcome ("实现智能化", "提升效率") is WORSE than no outcome.

5. Reverse Audit / 反向审计

After generating the tailored resume, simulate a senior interviewer reviewing it. Flag AI-generated feel, logical gaps, scope inflation, and buzzword defense risks. For every 🔴 Major issue, generate mock interview questions and STAR preparation sheets.

---

Workflow

Phase 1: Contextual Intelligence / 需求与背景深度挖掘

Philosophy / 理念: Understand not just what the JD says, but the market context, role level, and cultural norms that shape what the hiring manager actually looks for.

Step 1a: Accept JD + Market Research / 接收 JD + 市场调研

1. Accept JD input — URL or plain text.

   • If URL: use web_fetch to extract job description content.
   • If text: use directly.

2. Ask for company and role context / 收集公司+岗位信息:

   "请问这是哪个公司的岗位？具体岗位名称是什么？"

   • If the user provides company name and role:

3. Web search for market expectations / 岗位偏好搜索:

Use web_search with 2 targeted queries:

• Query 1: "2026 [company] [role] 招聘偏好 技能要求"
• Query 2: "[company] [role] 面试重点 产品文化"
• Fallback (if no company): "2026 [role] 岗位趋势 技能要求"

Present findings briefly: "🔍 市场调研: [2-3 key insights]"

These insights feed into keyword extraction and checkpoint design.

Step 1b: Unified Context Extraction / 统一上下文特征提取

Execute script extraction and LLM semantic analysis in one pass. Present a single consolidated table to the user.

Step 1b-i: Run script for hard requirements:

python scripts/jd_parser.py --file <jd_path> --json [--resume <resume_path>]

Script extracts: experience years, degree requirements, language tests, certifications, GPA, work authorization, date ranges.
Script does NOT extract: soft skills, domain knowledge, business capabilities → LLM handles these.

Position weight signal: Requirements in the first 20% of JD → likely Dealbreakers. Last 20% → likely Nice-to-have.

Step 1b-ii: Role Context Detection / 岗位上下文检测:

1. Role Level / 岗位层级:
   "应届"/"intern" → Campus | "3+ years"/"manager" → Social | "VP"/"总监" → Executive

2. Target Region / 目标区域:
   North America | UK/Ireland | DACH | East Asia | Middle East | Global/Remote

3. Document Type: CV (Academic/UK/EU) vs Resume (Industry/US)

4. Work Arrangement: Remote | Hybrid | On-site | Timezone-restricted

Output: "检测结果：Campus 岗位、东亚地区、Resume 格式、On-site。"

Step 1b-iii: LLM Keyword Extraction + Skill Clustering / 语义关键词提取 + 能力簇聚类:

Read full JD text + script JSON output. Extract keywords by category (Hard Skills, Soft Skills, Domain Knowledge, Experience Context), assign weights (🔴 Core / 🟡 Nice-to-have), then cluster into capability groups.

Output — Single Consolidated Table:

## 上下文特征提取 / Context Extraction

### 角色定位
- 岗位层级: [Campus/Social/Executive]
- 目标区域: [Region]
- 文档类型: [CV/Resume]
- 工作安排: [Remote/Hybrid/On-site]

### 关键词与能力簇
| # | 关键词 | 能力簇 | 权重 | JD 证据 |
|---|--------|--------|------|---------|

### 硬指标检查 (脚本)
| 指标 | JD 要求 | 简历状态 |
|------|---------|---------|

### 格式与兼容性风险
| 风险项 | 严重度 | 建议 |
|--------|--------|------|

Integrated Format & Risk Warnings / 整合格式与风险提示:

After reading the resume (Step 2a), consolidate ALL risk warnings into a single table above:

• ATS risk: Table-heavy layout, missing section headings, fragmented content from graphic PDF
• PDF reading fallback: If pdfplumber unavailable, note degraded quality
• Timeline gaps: ≥3 months between entries
• Portfolio links: Missing GitHub/Behance/etc. for relevant role types
• Tech stack freshness: Legacy-only tech in a category

Do NOT scatter these warnings across multiple phases — present them once here.

---

Phase 2: Strict Matching / 严谨匹配与风险识别

Philosophy / 理念: Go beyond literal keyword matching. Use semantic understanding and capability clusters to identify implicit matches, while grading confidence to prevent over-interpretation.

Step 2a: Read Resume with Structure Awareness / 结构感知读取

If the source resume is .docx, use the structured reader:

python scripts/main.py read-structured --resume <resume_path>

Structure Fallback / 结构降级兜底:

If Normal style > 80% of paragraphs → launch LLM Semantic Structure Recovery.
After recovery, present section boundaries to user for confirmation (see interaction_checkpoints.md — Structure Recovery section). Do NOT proceed without user confirmation.

Step 2b: Semantic Matching / 语义匹配

Compare Phase 1 keyword table against resume content.

Match levels:

Level	Definition	Default Confidence
✅ Direct	Keyword appears verbatim	High
🔄 Implicit	Same capability, different wording	High / Medium / Low
⚪ Gap	No related content found	—

Implicit match rules:

• MUST provide "Match Reason" column
• MUST assign confidence: High / Medium / Low
• Low confidence → forwarded to CP4 for user verification
• Anti-hallucination: If unsure, mark as Gap — do NOT inflate the score

Equivalence & Proxy Matching:

• JD asks CET-6, resume has IELTS 7.5 → ✅ Implicit (High)
• JD asks IELTS 7.0, resume has CET-4 → ⚪ Gap
• Cross-reference: IELTS 7.0 ≈ TOEFL 95-100 ≈ CET-6 550+ (mark as Implicit Medium when crossing test types)

Cultural Translation of Soft Skills:

• "抗压能力强" (CN) ≈ "Thrives in fast-paced environments" (US)
• Do NOT translate word-for-word — translate the underlying capability
• Offer as CP4 suggestions, never auto-apply

Match Score:

Score = (Direct×1.0 + Implicit_High×0.8 + Implicit_Med×0.5 + Implicit_Low×0.2 + Gap×0) / Total × 100%

Output format:

## 匹配度分析

### 概览
Match Score: XX% (Direct: X | Implicit High: X | Implicit Med: X | Gap: X)

### ✅ Direct Matches (X)
| JD Keyword | Resume Evidence | Section |

### 🔄 Implicit Matches (X)
| JD Keyword | Match Reason | Confidence | Resume Evidence |

### ⚪ Gaps (X)
| JD Keyword | Weight | Notes |

### ⚠️ Hard Requirement Alerts
| Requirement | JD Value | Resume Status | Action |

Language determination: Tailored resume language = JD language, NOT source resume language. Translate + culturally adapt if mismatch.

---

Phase 3: Dynamic Interaction / 动态交互调整

Philosophy / 理念: Not all resumes need the same level of adjustment. Route intelligently, then execute checkpoints interactively.

Step 3a: Three-Tier Routing / 三级路由

Hard Requirement Pre-check:

if script hard_gaps contain "dealbreaker" items (e.g., work authorization):
  → Alert: "⚠️ JD requires [X] which was not found. This is typically non-negotiable.
   Continue? (Yes/No)"

Match Score Routing:

Tier	Score	Flow
Tier 1: Fast-Track	≥ 90%	Generate draft + suggestion table directly, skip CP2/CP5
Tier 2: Full Flow	50%–89%	Execute CP1 → CP2 → CP3 → CP4 (with CP5 optional)
Tier 3: Alignment Check	< 50%	Confirm delivery intent before proceeding

Fast-Track Mode (Tier 1):

1. Generate "Recommended Adjustments" table + Markdown draft in ONE message
2. User confirms → proceed to Phase 4
3. User requests changes → apply and re-present
4. User switches to Full Mode → enter Tier 2 flow

Alignment Check (Tier 3):

"Match Score: [X]%. 简历与 JD 匹配度较低，请先确认：
1. 你确实想投这个方向的岗位吗？
2. 有没有简历里没写的相关经历？
3. 这个 JD 是参考还是实际投递目标？"

Memory Pre-fill: Before executing any checkpoint, check {resume_dir}/history/user_profile_cache.json for previously provided data. Pre-fill suggestions when keyword overlap is detected.

Step 3b: Execute Checkpoints / 执行介入点

CRITICAL: Follow the checkpoints in references/interaction_checkpoints.md. This file contains the Global Interaction Principle and detailed guidance for each checkpoint.

⚠️ GLOBAL HARD RULES / 全局硬约束:

1. EXPERIENCE ORDER HARD RULE:
   Work/internship entries MUST remain in REVERSE CHRONOLOGICAL ORDER.
   ONLY operation allowed: INCLUSION or EXCLUSION (取舍).
   Exception: Personal projects CAN be ordered by relevance.

2. PAGE LIMIT HARD RULE:
   Ask on first run: "简历需要控制在几页以内？(默认 1 页)"
   HARD constraint — final output MUST fit within the limit.
   If user requests more entries than budget allows → warn with specific numbers.

3. SPACE BUDGET CALCULATION:
   After CP1, estimate: Header ~2 lines, Education ~3-4 lines,
   Per work entry ~4-6 lines, Per project ~3-4 lines, Skills ~2-3 lines.
   1-page A4 budget: ~45-50 lines (11pt font).
   If total > budget → warn BEFORE generating.

Checkpoint execution order:

1. CP1: Experience Selection / 经历取舍 — Always execute.
   Present in reverse chronological order with ✅/⚠️/❌ recommendations.
   User confirms which to keep/remove.

2. CP2: Content Gaps / 内容缺口 — Execute in Tier 2.
   For each gap, use scenario-based prompts (not closed-ended questions).

3. CP3: Quantification / 量化补充 — Always execute.
   Apply Anti-Filler Rule strictly: progressive probing → fallback to original wording.
   Never use vague descriptors ("实现智能化", "提升效率") as substitutes.

4. CP4: Wording Upgrade / 措辞升级 — Always execute.
   Weak verb → Strong verb + Business impact.
   Apply cultural tone slider for target region.
   Includes Low-confidence implicit matches from Phase 2 for user verification.
   Every change must be confirmed by user (⚠️ for potential over-claims).

Multi-Currency Value Alignment: When resume uses CNY but JD targets another region, suggest approximate conversion with context. Use target currency as primary, original as parenthetical. Never fabricate numbers.

---

Phase 4: Delivery & Reverse Audit / 交付与反向审计

Philosophy / 理念: The tailored resume is not complete until it passes the "senior interviewer" test. Every claim must be defensible in an interview.

Step 4a: Generate & Deliver / 生成与交付

1. Generate Markdown draft — Based on all confirmed adjustments.
2. Final confirmation — Present complete Markdown to user for review.
   • User confirms → proceed to .docx generation.
   • User requests changes → apply and re-present.
3. Generate .docx — Use python-docx to create a styled .docx.
4. Optional: PDF — Use pandoc to convert .docx to PDF (if installed).

Save intermediate Markdown to history/YYYY-MM-DD_{company}_{role}.md. This is a deliverable, not a temp file — required for Phase 4b diff audit.

Shadow Resume (on-demand): If JD language differs from source resume language, offer to generate a parallel translation copy. Only generate if user explicitly requests.

Step 4b: Compliance Audit / 地域合规性审计

Based on Target Region from Phase 1, check for regional compliance violations:

IF North America:
  ❌ FORBIDDEN: Photo, age, gender, marital status, religion, nationality, salary history
  ⚠️ CAUTION: Home address (city/state only), graduation year

IF UK/Ireland:
  ❌ FORBIDDEN: Photo, age, marital status, gender, NI number

IF DACH:
  ✅ ACCEPTABLE (changing norms): Photo, DOB, marital status
  ❌ FORBIDDEN: Religion, political affiliation

IF East Asia:
  Varies by country — CN (photo optional in tech), JP (photo expected), KR (photo expected), SG (optional)

IF Global/Remote: Default to North America norms (most restrictive)

Output: Table with item, status, region rule, and action. 🔴 Critical violations require explicit user acknowledgment.

Step 4c: Reverse Audit / 反向审计

Step 4c-i: Interviewer Persona / 面试官画像:

Based on JD, construct a senior interviewer persona (role, seniority, industry, region). Infer evaluation priorities by role type:

• Product director → user growth attribution, cross-functional leadership, data-driven decisions
• Tech architect → technical 'why', trade-off decisions
• Finance director → numbers accuracy, risk awareness
• Junior role → learning trajectory, potential, cultural fit

Output: 👤 [Persona summary in 1-2 sentences]

Step 4c-ii: Persona-Based Review / 画像驱动审视:

Review the tailored resume through the interviewer's eyes. Flag:

#	Category	What to check
1	🤖 AI feel	Overly polished phrasing, buzzword stacking, formulaic structure
2	🕳️ Logical gaps	Outcomes without process, unexplained transitions
3	📏 Scope inflation	"Led" when actual role was "participated", exaggerated numbers
4	🎯 Buzzword defense	Terms candidate can't explain in interview
5	🌍 Cultural tone	Humble language for NA, aggressive language for East Asia

Output table with location, issue, severity (🟡/🔴), and suggestion.

Step 4c-iii: Interview Readiness / 面试备战:

For every 🔴 Major issue:

Mock Questions: Generate specific, challenging follow-up questions the interviewer might ask.

STAR Preparation Sheets: For each significantly modified bullet:

📌 "[exact bullet text]"

**S — Situation**: Business context, team/product/period
**T — Task**: Specific problem, goal or KPI
**A — Action**: Tools/methods, YOUR personal contribution
**R — Result**: How measured, baseline, time period

**Follow-up Risk**: "What would you do differently?" / "Biggest challenge?"

Rules:

• Only generate for 🔴 Major issues
• Questions must be specific, not generic
• Language matches JD language
• This output is SAVED in the audit log

Step 4d: Generate Audit Log / 生成审计日志

1. Diff audit — Compare source vs tailored:

   python scripts/main.py diff --source-docx <source_resume.docx> --tailored <history/YYYY-MM-DD_company_role.md> --company <company> --role <role> --json

2. ATS compatibility check:

   python scripts/ats_checker.py --resume <tailored_md> --keywords <comma_separated_jd_keywords> --region <target_region> --json

3. Compile final audit report combining:

   • Change summary (from diff_audit.py)
   • Risk classification (from Phase 3 CP4)
   • Compliance audit (from Step 4b)
   • Reverse audit (from Step 4c)
   • ATS check results

4. Save audit log alongside the tailored version.

Output Structure

{resume_directory}/
├── resume_source.docx          # User's original — NEVER modify
└── history/
    ├── 2026-04-14_腾讯_产品策划.md           # Tailored Markdown
    ├── 2026-04-14_腾讯_产品策划.docx          # Final deliverable
    ├── 2026-04-14_腾讯_产品策划_audit.md       # Audit log
    └── ...

Error Handling

• If no resume file is found: ask user to provide the path.
• If JD URL fails to fetch: ask user to paste JD text directly.
• If pandoc is not installed: install it before PDF conversion, or output .docx only.
• If python-docx is not installed: install it before .docx generation.
• If pdfplumber is not installed: PDF reader falls back to PyPDF2 or pdftotext. Warn user about potential quality loss. Do NOT abort.
• If JSON validation fails (jd_parser.py): retry once, then degrade to raw text with warning. Do NOT abort.

Dependencies

• python-docx — Read and write .docx files
• pdfplumber — Read PDF files (with PyPDF2 and pdftotext as fallbacks)
• pandoc — .docx to PDF conversion (optional — can also output .docx directly)