English | 中文
An open-source, offline-first AI skill that turns your recorded job interviews into a professional HTML debrief report — scored across 8 dimensions, with filler-word statistics, a segment-by-segment replay guide (with the interviewer's questions, your actual answers, and 60-second expert model answers), an industry perspective, and a concrete prepare-next-time plan.
Works with Claude Code (primary), and is compatible with Cursor and OpenAI Codex CLI via AGENTS.md.
No API key purchase is required to share or use this skill. It is just files — whoever runs it uses their own AI coding tool (Claude Code / Cursor / Codex) and their own subscription. Nothing is ever uploaded by the skill itself.
A single self-contained HTML report (dark, responsive, no external assets) with a sticky index:
| # | Section | What it does |
|---|---|---|
| 00 | Overview & verdict | Overall score gauge, strongest / costliest moment, key theme |
| 01 | 8-dimension scorecard | Tech match, depth, structure, impact, ownership, growth, motivation, communication (1–5 each) |
| 02–09 | Per-dimension deep dives | ✅ wins / |
| 10 | Fluency analysis | Objective um/uh/stutter counts + per-minute density + practice targets |
| 11 | Timeline | Key moments, color-coded good / neutral / bad |
| 12 | Segment-by-segment replay | The centerpiece: for every segment — the interviewer's original question, your actual answer (key points + verbatim), what went wrong, how to improve, and on heavy-loss questions a 60-second expert model answer you can rehearse |
| 13 | Industry perspective | Where this role & your profile sit in the current market |
| 14–16 | Prepare / tactics / reverse questions | Story bank (STAR-L), quantified-wins table, in-the-moment tactics, tailored questions to ask back |
| 17 | One-line verdict | Top-3 priority actions |
Dark by default, with a day mode. A ☀️/🌙 toggle in the top-right switches between the dark palette and a light one, and remembers your choice.
If you hand the skill the audio/video alongside the transcript, the report ships wired to it:
- Every timestamp becomes a ▶ play button — click any moment in the timeline, segment guide, or a quoted line and the built-in player jumps straight there. Reading "you hedged here" is one thing; hearing it is another.
- A live caption track with an on/off toggle follows the audio as it plays, so you can read and listen at once. Captions are embedded in the HTML, so it all works offline by double-clicking the file.
- Output is a folder holding the recording, the transcript, and the report together — share the folder and everything keeps working.
- Captions use the classic subtitle look — yellow with a black outline — so they stay readable over either theme.
| Input | Required? | Notes |
|---|---|---|
| CV / résumé | ✅ Required | PDF is fine. Anchors fit scoring + the story bank |
| Interview record | ✅ Required (one of two) | a) transcript .srt / .vtt / .txt (timestamps preferred), or b) audio/video — transcribed locally (e.g. whisper), never uploaded |
| JD (job description) | ⭐ Strongly recommended | Enables role-tailored fit analysis; report still works without it |
git clone https://github.com/GabrielPower1969/InterviewAnalysis.git
cp -r InterviewAnalysis/skills/interview-analysis ~/.claude/skills/Then in any Claude Code session, just share your CV + transcript (or audio) and say "analyze this interview" — the skill triggers automatically. Or invoke explicitly: /interview-analysis.
Add a project rule pointing at the skill instructions (Cursor reads AGENTS.md automatically in recent versions; otherwise):
Settings → Rules → Add rule → "For interview analysis tasks, follow the
instructions in skills/interview-analysis/SKILL.md of this repo."
Codex reads the repo's AGENTS.md automatically. Clone the repo, put your CV / JD / transcript inside (or reference absolute paths), and ask for an interview analysis.
- Offline-first: reading files, filler statistics (
scripts/filler_stats.py, pure local Python), and report generation all happen locally. - The skill instructs the AI to never upload your CV / transcript / media to any external service, and to keep personal contact details out of the report.
- Your interview data stays in your folders; only the skill files live in this repo. No sample interviews, CVs, or personal data are included.
skills/interview-analysis/
├── SKILL.md # The skill: workflow, inputs, report structure
├── references/rubric.md # 8-dimension scoring rubric (1–5 levels, weights)
├── assets/report_template.html # Self-contained HTML report template (+ audio player & captions)
└── scripts/
├── filler_stats.py # Offline filler/disfluency counter (srt/vtt/txt)
└── srt_to_cues.py # Transcript -> inline caption cues for the report
MIT — see LICENSE.
一个开源、离线优先的 AI 技能：把你的面试录音变成一份专业的 HTML 复盘报告——八维评分、吞吐词统计、逐段精读（含面试官原问、你的原答、以及可直接照着练的 60 秒专家范例回答）、行业视角、以及下次面试的备战清单。
以 Claude Code 为主，通过 AGENTS.md 兼容 Cursor 和 OpenAI Codex CLI。
分享或使用本技能无需购买 API key——它只是一组文件，谁使用就用谁自己的 AI 编程工具（Claude Code / Cursor / Codex）和订阅额度。技能本身不会上传任何数据。
单个自包含 HTML 文件（深色主题、响应式、无外部依赖），带固定目录索引：
| # | 版块 | 说明 |
|---|---|---|
| 00 | 总览与结论 | 综合评分仪表盘、最强/最伤时刻、关键主题 |
| 01 | 八维评分卡 | 技术匹配、深度、表达结构、业务影响、协作、成长、动机、沟通（各 1–5 分） |
| 02–09 | 逐维度详评 | ✅ 做得好 / |
| 10 | 吞吐词分析 | um/uh/结巴客观计数 + 每分钟密度 + 可量化练习目标 |
| 11 | 时间线速览 | 高光/中性/失分关键时刻 |
| 12 | 逐段精读 | 核心版块：每段给出面试官原问、你的回答重点（含原话）、问题在哪、怎么提高；强烈失分题附🎯 60 秒专家范例回答，可对着音频逐段复盘、照着练 |
| 13 | 行业视角 | 这个岗位和你的画像在当前市场的位置 |
| 14–16 | 备战清单 / 临场战术 / 反问银行 | STAR-L 故事库、量化战绩表、临场行为修正、定制反问 |
| 17 | 一句话总结 | 三条最高优先级行动 |
默认夜间模式，另配白天模式。 右上角 ☀️/🌙 按钮一键切换深浅配色，选择会被记住。
把录音和转录一起交给技能，报告会直接和音频挂上钩：
- 每个时间戳都是一个 ▶ 播放按钮——时间线、逐段精读、引用原话，点哪里就从哪里开始播。看到「你这里犹豫了」和亲耳听到，感受完全不同。
- 实时字幕 + 开关，跟着音频滚动，可边听边读；字幕内嵌在 HTML 里，双击打开即可离线使用。
- 输出是一个文件夹，录音、转录、报告三件套放在一起——整个文件夹分享出去，功能照常。
- 字幕采用经典电视字幕样式——外黑里黄（黄字黑描边），在深色和浅色背景下都清晰。
| 输入 | 是否必需 | 说明 |
|---|---|---|
| 简历 CV | ✅ 必选 | PDF 即可，用于匹配度评分和故事库挖掘 |
| 面试记录 | ✅ 必选（二选一） | a) 文字转录 .srt / .vtt / .txt（带时间戳最佳）；或 b) 音频/视频——用本地工具（如 whisper）转录，绝不上传 |
| JD 职位描述 | ⭐ 强烈推荐 | 有 JD 可做针对岗位的匹配分析；没有也能出完整报告 |
git clone https://github.com/GabrielPower1969/InterviewAnalysis.git
cp -r InterviewAnalysis/skills/interview-analysis ~/.claude/skills/之后在任意 Claude Code 会话中，把简历 + 转录（或音频）发给它，说一句**「分析这个面试」**即可自动触发；也可以显式调用 /interview-analysis。
添加一条项目规则指向技能说明（新版 Cursor 会自动读取 AGENTS.md；否则）：
设置 → Rules → 新建规则 → 「面试分析任务请遵循本仓库
skills/interview-analysis/SKILL.md 中的说明。」
Codex 会自动读取仓库的 AGENTS.md。克隆仓库后，把简历 / JD / 转录放进来（或引用绝对路径），然后请求做面试分析即可。
- 离线优先：读文件、吞吐词统计（
scripts/filler_stats.py，纯本地 Python）、报告生成全部在本地完成。 - 技能明确要求 AI 不得把简历/转录/音视频上传到任何外部服务，且报告中不包含电话、邮箱等个人联系方式。
- 你的面试数据留在你自己的文件夹里；本仓库只包含技能文件，不含任何示例面试、简历或个人数据。
MIT — 见 LICENSE。