Resume JD Matcher

The last check before you hit submit. ATS and HR filter on keywords and hard requirements first. Matching is not fabrication — it is translating things you genuinely did into the employer's vocabulary.

Distinct from job-application-screener (which decides whether to apply across a batch) and vibecoding-resume-story (which turns one project into a resume entry).

Three steps

1. Machine check (always run)

python "<this skill dir>/scripts/match_resume.py" <JD.txt> <resume.txt>

Outputs JSON with three blocks:

• hard_requirements — degree, years, certificates, location; each judged 达标 / 缺失 / 未提及
• keyword_coverage — which JD skill/tool/domain terms appear in the resume, coverage %, missing list
• duplicate_lines — same sentence repeated (copy-paste residue)

2. Semantic match (AI layer, covers the script's blind spots)

1. Synonyms — JD says 用户增长, resume says 拉新. No literal hit but a real hit. Label these 语义匹配 and list them separately from true gaps.
2. Hidden signals — translate euphemisms. 抗压能力强 = heavy overtime. 扁平化管理 = nobody will mentor you. State the implication plainly.
3. Placement — is the most relevant experience on the first screen? Is irrelevant experience eating too much space?

3. Report

【结论】硬条件 X/Y 达标 | 关键词覆盖 62% | 综合：可投(需改) / 勉强 / 不建议投
【硬条件明细】学历✓ 年限✗(JD要3年,简历可算出2年4个月,建议写"近3年") 证书?
【缺失关键词】分三档：
    真没做过      → 不建议编
    做过没写      → 给改写句
    语义匹配      → 建议换词
【改写示例】JD原词 → 简历改写句（只基于用户提供的真实经历）
【工作信号】"抗压能力强"+"弹性工作制" → 大概率经常加班，请确认能否接受

Step 4 (optional) — Rewrite table, when the ask is "照着 JD 改简历"

"匹配吗" ends at the report. "帮我改" needs something the user can paste into their own file. Give one row per location:

Character-count discipline

去空白后计数。分级：|Δ| ≤ 3 绿（版面不动）、≤ 10 黄、> 10 红， 并且每个红的都要在同行附一个压缩版。不要为了对齐字数硬凑——报出差值，给短版本。

Real case (2026-10-21): 26 处，20 处在 ±5 内。6 个超差的都是故意的 （一个 1 行的项目升格成 4 行、新增一条技能行、一节压缩腾版面）， 每行都写了「版面还超的话先删哪几个字」。

优先原地替换，不要推倒重写

用户原本的措辞是底稿不是废纸。改「素材整理」→「交付标准」这类换词， 比整段重写安全得多，也更容易被他接受。

主动找被错词描述的项目

改之前先扫一遍他的仓库/项目列表，找已经是目标岗位形状、但用了旧领域词汇描述的东西。 这是改写里杠杆最高的一处，而且不需要编任何新东西。

Real case： ad-compliance-checker 在简历上写的是「带货文案自动三级风险分级」——业务语言， 1 行。它实际是规则引擎 + 分级用例集 + 回归集。改成测试语言后升格为 4 行， 成了整份简历里唯一能正面回答「你做过什么测试」的东西。

Red lines

• 🔴 Never suggest inventing experience. For keywords the candidate genuinely lacks, write 不建议编. Offer a real adjacent experience instead, or suggest addressing it at interview.
• 🔴 Years are computed from the resume's own dates. Never round up on the candidate's behalf.
• 🔴 If a hard requirement fails, say 不建议投. No consolation padding.
• 🔴 Remind the user to desensitize — no ID numbers, home address, or photos in files handed to analysis.

Example

帮我看看这个岗投得了吗？（贴 JD + 简历）

Machine check reports 年限 gap (JD wants 3 years, resume computes to 2y4m) and 68% keyword coverage. AI layer flags 「私域运营」as a true gap (never done → 不建议编), 「内容矩阵」as done-but-unwritten (rewrite from the 小红书 account work already in the resume), and 「用户增长」as a synonym of 「拉新」. Verdict: 可投(需改).