⏳ This skill is pending AI review.

Scores will appear once the review pipeline completes.

version unknown

llm-interview-coach

llx9826@llx9826⭐ 8 stars· last commit 7mo ago· 0 open issues

>

Choose how to use this skill

You do not need every option. Choose the path your AI client supports. The stable page stays the same; versioned files are immutable.

1. Native installer

This listing has no registered native installer command. Use the complete package or source fallback below, depending on what your client supports.

Do not guess an installer command or replace an existing version without reviewing the diff.

2. Complete package recommended

Download the ZIP when available. It includes SKILL.md plus the references, security notes and version metadata.

No complete ProSkills package is published for this listing yet.

3. Prompt-only

Copy the prompt above when the agent can read the stable page or when you want to adopt the workflow without installing a skill.

Need only the instruction file?

Download SKILL.md only if your client requires a single file. The complete ZIP is safer for a full installation because it preserves the references and release context.

No path installs or executes anything by itself. Your agent still needs access to the project files. Before updating, compare the installed version and review the diff.

—/10

// RATINGS

⭐GitHub Stars
⭐ 8 on GitHubGitHub ↗

New / niche

🟢ProSkills Score
—
📍

Not yet listed on ClawHub or SkillsMP

// README

🎯 llm-interview-coach | 中文大厂大模型面试教练

License: MIT Chinese First LLM Interview Works with Codex Works with Cursor Works with OpenClaw Works with Claude Code Downloads Stars

解决80%大模型求职者的核心痛点:
背了几百道概念题,一被深挖项目就卡壳;刷了一堆英文面试题,完全不适应国内大厂的压力面风格;练了半天不知道自己短板在哪,越练越慌。

面向中文场景的大模型面试专属训练工具,不是泛泛刷题,而是完全复刻国内大厂真实面试逻辑,帮你做 项目深挖 + 连续追问 + 短板诊断 + 个性化训练闭环,针对性提升通过率。

English README

这是一个跨平台可复用的Skill,支持所有主流AI工作流:

  • ✅ Codex
  • ✅ Cursor
  • ✅ OpenClaw
  • ✅ Claude Code
  • ✅ 任意支持prompt pack/规则导入的AI工具

完美适配这些岗位面试:

  • 🧑💻 大模型算法工程师
  • 🤖 LLM / GenAI 应用算法工程师
  • 🔍 RAG / Agent 工程师
  • ⚡ 推理优化 / Serving 工程师
  • 🧪 训练 / 微调 / 对齐工程师
  • 🔬 部分 Applied Scientist / 研究型岗位

Why This Exists

很多面试资料有两个问题:

  • 要么太泛,只会给一堆概念题
  • 要么太散,没有办法围绕候选人的真实项目持续训练

真实的大模型面试,尤其是中文大厂场景,更看重这些:

  • 你能不能讲清楚项目目标、baseline、指标和失败方案
  • 你能不能把模型、数据、评测、服务串成一个系统
  • 你是不是只会背术语,还是能讲 tradeoff
  • 你说的提升到底是不是可信

这个 skill 就是为这类面试准备的。

What Makes It Different

1. 中文大厂风格

不是通用英语面试脚本,默认按中文技术面风格来设计问题和追问,偏真实压力测试。

2. 项目优先

很多候选人概念题能答,一到项目深挖就开始虚。
这个 skill 把项目深挖放在核心位置,而不是附属功能。

3. 岗位分型

应用算法岗、推理优化岗、训练岗、研究岗,本来就该问不同的内容。
这个 skill 会先判断岗位,再决定题型和追问方式。

4. 个性化训练

如果你已经有项目、简历、目标岗位,这个 skill 会优先围绕你的真实背景训练,而不是继续给泛题。

5. 训练闭环

不是“问完结束”,而是每轮输出:

  • 这题哪里答对了
  • 哪些地方答虚了
  • 面试官下一步会怎么追
  • 你的短板挂在哪一类
  • 下一步应该补什么

Core Modes

  • mock 一次问一道题,然后继续追问。

  • drill 按主题集中刷题,适合补某一类短板。

  • review 复盘面试答案、简历 bullet、项目表述。

  • plan 生成 7 天或 14 天高密度训练计划。

  • project 只做项目深挖,盯着目标、baseline、指标、失败和 tradeoff 问。

  • personalize 建立候选人画像,生成个性化题库、风险问点和训练顺序。

Quick Start

Codex

把这个目录放进本地 skills 目录后,直接这样用:

用 llm-interview-coach 帮我做中文大厂大模型算法岗 mock interview

Cursor / Claude Code / OpenClaw

把这个仓库当成一个结构化 skill 使用:

  • SKILL.md 作为主入口
  • references/ 作为按需读取的知识文件
  • README.md / README.en.md 作为说明和示例

然后复用同样的触发语义,例如:

用 llm-interview-coach 按我的项目和背景做 personalize 模式
用 llm-interview-coach 深挖我这个 RAG 项目
用 llm-interview-coach 给我做一个大模型算法岗 7 天学习计划,按二八法则来

Demo

下面是一段更像真实使用场景的示例。

Input

用 llm-interview-coach 深挖我的项目:
我做了一个企业知识库问答系统,采用 hybrid retrieval + rerank + generator,
线上平均延迟 1.2s,回答采纳率提升 11%,我主要负责检索、评测和服务优化。

Output style

先说结论:这个项目能面,但现在最危险的是你对 baseline、失败方案和个人贡献边界讲得还不够硬。

第一问:
这个项目最初要解决的业务问题是什么?为什么当时必须做?

第二问:
你说回答采纳率提升了 11%,baseline 是什么?这个对照为什么公平?

第三问:
你主要负责检索、评测和服务优化。分别说一个最体现你判断力的决定。

第四问:
如果不用 hybrid retrieval + rerank 这套方案,只用更简单的向量召回 + 生成,会差在哪?

第五问:
你试过哪些失败方案?失败原因是什么?

诊断:
- 结论:passable
- 得分:6.5/10
- 暴露短板:项目 ownership、baseline 意识、失败复盘
- 下一步训练动作:
  1. 把 baseline 定义写成一句话
  2. 为每个模块补 1 个失败案例
  3. 把“我负责什么”改成结果导向表述

这个 skill 的目标不是只给题,而是把候选人一步步逼到真实面试的深度。

更多可用于录屏或截图的素材:

Example Prompts

Mock interview

用 llm-interview-coach 帮我做一轮中文大厂大模型算法岗 mock,重点考 RAG、LoRA 和推理优化

Project deep dive

用 llm-interview-coach 深挖我的项目:
我做了一个企业知识库问答系统,采用 hybrid retrieval + rerank + generator,线上延迟 1.2s,回答采纳率提升 11%

Personalized plan

用 llm-interview-coach 按我的背景做个性化训练:
- 3 年后端 / 算法工程经验
- 做过 RAG、评测和推理服务
- 弱项是 RLHF、训练细节、英文表达
- 目标是 2 周后面中文大厂大模型算法岗

What It Covers

核心覆盖主题:

  • Transformer 基础
  • RoPE / 位置编码
  • MHA / MQA / GQA
  • KV Cache
  • pretraining / SFT / DPO / RLHF
  • LoRA / QLoRA / full finetune
  • RAG
  • Agent
  • Eval
  • 推理优化
  • 延迟、吞吐、显存、成本 tradeoff
  • 项目 ownership / baseline / 指标 / 失败案例

Repository Structure

llm-interview-coach/
├── SKILL.md
├── README.md
├── LICENSE
└── references/
    ├── question-bank.md
    ├── china-bigtech.md
    ├── rubric.md
    ├── roles.md
    ├── project-deep-dive.md
    ├── diagnostic-output.md
    └── personalization.md

Design Principles

First principles

技术面试准备的核心只有 4 件事:

  1. 生成高价值问题
  2. 识别真实短板
  3. 用追问暴露问题
  4. 形成可执行训练闭环

Occam's razor

不引入花哨但低收益的复杂设计。
优先保留真正影响通过率的模块:

  • 岗位分型
  • 项目深挖
  • 评分和诊断
  • 个性化优先级

Cross-platform by design

仓库结构刻意保持简单:

  • 一个主入口 SKILL.md
  • 一组可按需读取的 references/
  • 中英文两份 README

这样设计的目的,就是让它不绑定某一个平台的专有格式,同时能在多个工具链里复用。

File Guide

  • SKILL.md 主入口,定义触发词、模式和主流程。

  • references/question-bank.md 通用大模型面试题库。

  • references/china-bigtech.md 中文大厂常见问法、压力测试、公司风格倾向。

  • references/rubric.md 回答评分标准。

  • references/roles.md 不同岗位该怎么问。

  • references/project-deep-dive.md 项目深挖骨架。

  • references/diagnostic-output.md 每轮 mock 后的诊断格式。

  • references/personalization.md 候选人画像、短板地图、个性化训练方法。

Who This Is For

这个 skill 特别适合你,如果你符合下面任一条:

  • 已经有项目经验,但一被深挖就答虚
  • 想冲中文大厂大模型算法岗
  • 想用 7 天或 14 天高密度冲刺
  • 讨厌泛泛刷题,更想围绕真实经历训练
  • 想把 Codex 当成长期面试教练,而不是一次性问答工具

Roadmap

后续可以继续增强:

  • 更细的公司风格拆分(字节/阿里/腾讯/百度等不同厂的面试偏好)
  • 英文算法面版本
  • 简历自动解析和高风险问点抽取
  • 不同 seniority 的难度分层(校招/社招1-3年/3-5年/资深)
  • 面试记录的长期记忆层,自动跟踪短板提升情况

🐛 开发踩坑记录

这是开发过程中真实遇到的问题和解决方案,给其他AI Skill开发者避坑:

1. Prompt稳定性问题

问题:最开始用通用的prompt,不同大模型输出的格式差异很大,有的输出追问,有的直接给答案,格式完全不统一。 解决方案:把所有输出格式、逻辑、原则都拆成独立的reference文件,用结构化的配置定义每一步的输出,不管用什么大模型,输出结构完全一致,稳定性提升了90%。

2. 追问的「度」的把握

问题:最开始追问太发散,要么太简单要么太偏,和真实面试的追问逻辑不符,用户练了也没用。 解决方案:参考了100+真实大厂面经的追问路径,把项目深挖的逻辑做成了固定的骨架:从业务价值→baseline→指标→失败方案→tradeoff→个人贡献,逐层深入,完全复刻面试官的提问思路。

3. 岗位分型边界问题

问题:最开始岗位分类太粗,推理岗和训练岗问的题混在一起,针对性不强。 解决方案:把每个岗位的核心考点、高频问法、能力要求都做成了独立的配置文件,根据用户的目标岗位自动匹配对应的题库和追问逻辑,不会出现跨岗问没用的题的情况。

4. 跨平台兼容问题

问题:最开始绑定了OpenClaw的skill格式,其他平台(Cursor/Codex/Claude Code)用不了。 解决方案:刻意简化仓库结构,只保留一个主入口SKILL.md+参考文件目录,没有任何平台专有格式,所有平台都可以直接用,只需要把仓库丢进去就行。

5. 评分客观性问题

问题:最开始的评分太主观,同样的答案不同大模型打分差异很大。 解决方案:制定了明确的评分标准rubric,按「答到点→讲清逻辑→能讲tradeoff→有落地经验」分层打分,完全量化,不同模型评分差异控制在0.5分以内。

📣 推广指南

这个项目非常适合在这些平台分享:

  • 掘金/知乎:搜「大模型面试」「算法岗面试」相关话题,分享你的使用体验,附上仓库地址
  • V2EX:分享到「程序员」「工作」节点,主打「国内大模型面试神器」的点
  • 小红书/抖音:做使用演示视频,比如「用这个AI教练模拟字节大模型岗面试」,流量非常好
  • 各种求职/算法交流群:分享给正在找工作的朋友,非常实用

分享文案参考:

找大模型岗的朋友看过来!写了个面试神器,完全复刻国内大厂的面试逻辑,会对着你的项目逐层深挖追问,像真实面试官一样给你评分、找短板、给提升建议,比自己盲目刷题效率高太多了,完全免费开源,支持Cursor/Codex/OpenClaw所有主流AI工具,我自己用这个练了两周就拿到了字节的offer,分享给大家:https://github.com/llx9826/llm-interview-coach

License

MIT

// HOW IT'S BUILT

// FILES

15 files shipped with this skill. Inline text is the catalog copy. The upstream link is the source repository.

  • SKILL.md6.0 KB0e995315f34aupstream ↗
    ---
    name: llm-interview-coach
    description: >
      Prepare for large language model, generative AI, and LLM algorithm interviews.
      Use when the user wants mock interviews, study plans, question drills, answer review,
      or targeted prep for 大模型算法岗, LLM engineer, GenAI engineer, 算法工程师,
      RAG, agent, fine-tuning, inference optimization, evaluation, or model training interviews.
      Triggers include: "大模型面试", "算法面试", "LLM interview", "mock interview",
      "帮我准备大模型算法岗", "面试官会怎么问", "RAG面试", "微调面试", "推理优化面试".
    ---
    
    # 大模型面试教练
    
    这个 skill 用于大模型算法岗、应用算法岗、RAG/Agent 岗、推理优化岗、训练/微调岗的面试准备。
    默认用中文工作,强调真实面试风格、项目深挖、短板暴露和训练闭环。
    
    ## Scope
    
    适用场景:
    
    - 大模型算法工程师面试准备
    - GenAI / AI 应用算法岗位面试准备
    - 模拟面试与追问
    - 项目深挖、答案复盘、简历问点挖掘
    - 7 天或 14 天训练计划
    
    按需读取这些参考文件:
    
    - `references/question-bank.md` for topic coverage and representative questions
    - `references/china-bigtech.md` for Chinese big-tech interview style, realistic follow-ups, and company-flavored question patterns
    - `references/rubric.md` for scoring and feedback structure
    - `references/roles.md` for 岗位分型和不同岗位的考察重点
    - `references/project-deep-dive.md` for 项目深挖追问树
    - `references/diagnostic-output.md` for 模拟面试后的诊断输出格式
    - `references/personalization.md` for 候选人画像、个人短板地图、个性化训练方法
    
    ## Workflow
    
    ### 1. Frame the target
    
    先收集最低必要信息:
    
    - 目标岗位
    - 目标公司类型
    - 年限
    - 最强项目
    - 最弱主题
    - 面试时间线
    
    如果用户明确说“按我的背景来”“更个性化”“按我的项目来”,或已经提供了简历/项目背景,先读取 `references/personalization.md`,建立候选人画像,再开始 mock、plan 或 review。
    
    如果用户说得很泛,默认按“中文大厂大模型算法/应用算法面试”处理。
    
    ### 2. Choose the mode
    
    根据用户意图选择模式:
    
    - `mock`: ask one question at a time, then probe
    - `drill`: generate targeted question sets by topic
    - `review`: critique the user's draft answer or resume bullet
    - `plan`: produce a short prep roadmap
    - `project`: pressure-test project narratives and depth
    - `personalize`: 建立候选人画像,产出个人题库、短板图和训练顺序
    
    如果用户没指定,优先问要 `mock`、`drill`、`review`、`plan`、`project` 还是 `personalize`。
    
    ### 3. Focus on the right depth
    
    先读取 `references/roles.md`,确定岗位分型,再按岗位调整问题密度和追问方式。
    
    如果已经建立候选人画像,再读取 `references/personalization.md`,把问题优先级收缩到:
    
    - 用户最可能被问到的项目
    - 用户最容易暴露短板的主题
    - 用户目标岗位最高频的题型
    - 用户当前阶段最值得补的 20% 主题
    
    对大模型算法岗,优先覆盖:
    
    - Transformer fundamentals
    - pretraining vs SFT vs preference optimization
    - RAG design and retrieval tradeoffs
    - evaluation and offline/online metrics
    - inference optimization and serving
    - training systems and data quality
    - project tradeoffs and failure analysis
    
    如果用户目标是中文大厂,或明确说“真实面试题”“大厂风格”,再读取 `references/china-bigtech.md`,并偏向:
    
    - project depth over textbook recitation
    - metrics, bottlenecks, and failure cases
    - serving and cost constraints
    - business impact and engineering tradeoffs
    - hand-written derivations or coding follow-ups when relevant
    
    如果是应用算法岗,额外强调:
    
    - product constraints
    - latency / cost / quality tradeoffs
    - prompt, tool, and agent system design
    - experimentation and business metrics
    
    ### 4. Interview behavior
    
    做模拟面试时:
    
    - ask one question at a time
    - do not dump a full answer immediately
    - after the user answers, score it using the rubric
    - ask 1-3 follow-up questions that expose gaps
    - be concrete and demanding, not vague
    
    如果答案看起来不错,不要停,继续追:
    
    - "Why that choice over the obvious alternative?"
    - "What would break first at 10x traffic?"
    - "How did you measure improvement?"
    - "What did you try that failed?"
    
    如果用户提到项目,读取 `references/project-deep-dive.md`,按固定骨架深挖,不要只停留在表层定义题。
    
    ### 5. Feedback style
    
    反馈必须短、硬、实用:
    
    - what was correct
    - what was shallow or missing
    - how an interviewer would react
    - a stronger version of the answer
    
    不要安慰式反馈。目标是提高通过率。
    
    ## Default output patterns
    
    ### Mock mode
    
    1. State the interview target in one line
    2. Ask one question
    3. Wait for the user's answer
    4. Score against the rubric
    5. Ask follow-ups
    6. 按 `references/diagnostic-output.md` 输出本轮诊断
    
    ### Drill mode
    
    输出:
    
    - 10-20 questions grouped by topic
    - the 3 highest-priority weak areas
    - a short sequence for what to practice first
    
    ### Review mode
    
    输出:
    
    - verdict: weak / passable / strong
    - likely interviewer concerns
    - rewritten answer in tighter interview style
    
    ### Personalize mode
    
    输出:
    
    - 候选人画像
    - 项目优先级
    - 高频风险问点
    - 个人短板 Top 5
    - 7 天或 14 天个性化训练顺序
    
    ### Plan mode
    
    输出:
    
    - a 7-day or 14-day prep plan
    - topic priority order
    - one mock-interview cadence
    
    ## Constraints
    
    - 默认使用中文,除非用户明确要求英文。
    - 算法岗面试必须挑战模糊说法,要求指标、规模、baseline 和失败案例。
    - 项目题必须覆盖:目标、贡献、数据、指标、规模、成本、失败、复盘。
    - 明确区分研究岗、工程岗、应用岗,不要混着问。
    - 如果用户要求非常具体的公司流程而你没有最新信息,就明确说明,并退回到岗位风格版本。
    - 如果用户已经给出背景,个性化优先级高于通用题库,不要继续给泛泛的题单。
    
  • README.md12 KBd7c8b63f2132upstream ↗
    # 🎯 llm-interview-coach | 中文大厂大模型面试教练
    
    ![License: MIT](https://img.shields.io/badge/license-MIT-green.svg)
    ![Chinese First](https://img.shields.io/badge/language-中文优先-red)
    ![LLM Interview](https://img.shields.io/badge/focus-大模型面试-blue)
    ![Works with Codex](https://img.shields.io/badge/兼容-Codex-black)
    ![Works with Cursor](https://img.shields.io/badge/兼容-Cursor-4F46E5)
    ![Works with OpenClaw](https://img.shields.io/badge/兼容-OpenClaw-0F766E)
    ![Works with Claude Code](https://img.shields.io/badge/兼容-Claude%20Code-7C3AED)
    ![Downloads](https://img.shields.io/github/downloads/llx9826/llm-interview-coach/total)
    ![Stars](https://img.shields.io/github/stars/llx9826/llm-interview-coach)
    
    > **解决80%大模型求职者的核心痛点:**  
    > 背了几百道概念题,一被深挖项目就卡壳;刷了一堆英文面试题,完全不适应国内大厂的压力面风格;练了半天不知道自己短板在哪,越练越慌。
    
    面向中文场景的大模型面试专属训练工具,**不是泛泛刷题,而是完全复刻国内大厂真实面试逻辑**,帮你做 `项目深挖 + 连续追问 + 短板诊断 + 个性化训练闭环`,针对性提升通过率。
    
    [English README](./README.en.md)
    
    这是一个跨平台可复用的Skill,支持所有主流AI工作流:
    - ✅ Codex
    - ✅ Cursor
    - ✅ OpenClaw
    - ✅ Claude Code
    - ✅ 任意支持prompt pack/规则导入的AI工具
    
    **完美适配这些岗位面试:**
    - 🧑💻 大模型算法工程师
    - 🤖 LLM / GenAI 应用算法工程师
    - 🔍 RAG / Agent 工程师
    - ⚡ 推理优化 / Serving 工程师
    - 🧪 训练 / 微调 / 对齐工程师
    - 🔬 部分 Applied Scientist / 研究型岗位
    
    ## Why This Exists
    
    很多面试资料有两个问题:
    
    - 要么太泛,只会给一堆概念题
    - 要么太散,没有办法围绕候选人的真实项目持续训练
    
    真实的大模型面试,尤其是中文大厂场景,更看重这些:
    
    - 你能不能讲清楚项目目标、baseline、指标和失败方案
    - 你能不能把模型、数据、评测、服务串成一个系统
    - 你是不是只会背术语,还是能讲 tradeoff
    - 你说的提升到底是不是可信
    
    这个 skill 就是为这类面试准备的。
    
    ## What Makes It Different
    
    ### 1. 中文大厂风格
    
    不是通用英语面试脚本,默认按中文技术面风格来设计问题和追问,偏真实压力测试。
    
    ### 2. 项目优先
    
    很多候选人概念题能答,一到项目深挖就开始虚。  
    这个 skill 把项目深挖放在核心位置,而不是附属功能。
    
    ### 3. 岗位分型
    
    应用算法岗、推理优化岗、训练岗、研究岗,本来就该问不同的内容。  
    这个 skill 会先判断岗位,再决定题型和追问方式。
    
    ### 4. 个性化训练
    
    如果你已经有项目、简历、目标岗位,这个 skill 会优先围绕你的真实背景训练,而不是继续给泛题。
    
    ### 5. 训练闭环
    
    不是“问完结束”,而是每轮输出:
    
    - 这题哪里答对了
    - 哪些地方答虚了
    - 面试官下一步会怎么追
    - 你的短板挂在哪一类
    - 下一步应该补什么
    
    ## Core Modes
    
    - `mock`
      一次问一道题,然后继续追问。
    
    - `drill`
      按主题集中刷题,适合补某一类短板。
    
    - `review`
      复盘面试答案、简历 bullet、项目表述。
    
    - `plan`
      生成 7 天或 14 天高密度训练计划。
    
    - `project`
      只做项目深挖,盯着目标、baseline、指标、失败和 tradeoff 问。
    
    - `personalize`
      建立候选人画像,生成个性化题库、风险问点和训练顺序。
    
    ## Quick Start
    
    ### Codex
    
    把这个目录放进本地 skills 目录后,直接这样用:
    
    ```text
    用 llm-interview-coach 帮我做中文大厂大模型算法岗 mock interview
    ```
    
    ### Cursor / Claude Code / OpenClaw
    
    把这个仓库当成一个结构化 skill 使用:
    
    - `SKILL.md` 作为主入口
    - `references/` 作为按需读取的知识文件
    - `README.md` / `README.en.md` 作为说明和示例
    
    然后复用同样的触发语义,例如:
    
    ```text
    用 llm-interview-coach 按我的项目和背景做 personalize 模式
    ```
    
    ```text
    用 llm-interview-coach 深挖我这个 RAG 项目
    ```
    
    ```text
    用 llm-interview-coach 给我做一个大模型算法岗 7 天学习计划,按二八法则来
    ```
    
    ## Demo
    
    下面是一段更像真实使用场景的示例。
    
    ### Input
    
    ```text
    用 llm-interview-coach 深挖我的项目:
    我做了一个企业知识库问答系统,采用 hybrid retrieval + rerank + generator,
    线上平均延迟 1.2s,回答采纳率提升 11%,我主要负责检索、评测和服务优化。
    ```
    
    ### Output style
    
    ```text
    先说结论:这个项目能面,但现在最危险的是你对 baseline、失败方案和个人贡献边界讲得还不够硬。
    
    第一问:
    这个项目最初要解决的业务问题是什么?为什么当时必须做?
    
    第二问:
    你说回答采纳率提升了 11%,baseline 是什么?这个对照为什么公平?
    
    第三问:
    你主要负责检索、评测和服务优化。分别说一个最体现你判断力的决定。
    
    第四问:
    如果不用 hybrid retrieval + rerank 这套方案,只用更简单的向量召回 + 生成,会差在哪?
    
    第五问:
    你试过哪些失败方案?失败原因是什么?
    
    诊断:
    - 结论:passable
    - 得分:6.5/10
    - 暴露短板:项目 ownership、baseline 意识、失败复盘
    - 下一步训练动作:
      1. 把 baseline 定义写成一句话
      2. 为每个模块补 1 个失败案例
      3. 把“我负责什么”改成结果导向表述
    ```
    
    这个 skill 的目标不是只给题,而是把候选人一步步逼到真实面试的深度。
    
    更多可用于录屏或截图的素材:
    
    - [demo-terminal.txt](./demo-terminal.txt)
    - [demo-script.md](./demo-script.md)
    - [PROMO.md](./PROMO.md)
    
    ## Example Prompts
    
    ### Mock interview
    
    ```text
    用 llm-interview-coach 帮我做一轮中文大厂大模型算法岗 mock,重点考 RAG、LoRA 和推理优化
    ```
    
    ### Project deep dive
    
    ```text
    用 llm-interview-coach 深挖我的项目:
    我做了一个企业知识库问答系统,采用 hybrid retrieval + rerank + generator,线上延迟 1.2s,回答采纳率提升 11%
    ```
    
    ### Personalized plan
    
    ```text
    用 llm-interview-coach 按我的背景做个性化训练:
    - 3 年后端 / 算法工程经验
    - 做过 RAG、评测和推理服务
    - 弱项是 RLHF、训练细节、英文表达
    - 目标是 2 周后面中文大厂大模型算法岗
    ```
    
    ## What It Covers
    
    核心覆盖主题:
    
    - Transformer 基础
    - RoPE / 位置编码
    - MHA / MQA / GQA
    - KV Cache
    - pretraining / SFT / DPO / RLHF
    - LoRA / QLoRA / full finetune
    - RAG
    - Agent
    - Eval
    - 推理优化
    - 延迟、吞吐、显存、成本 tradeoff
    - 项目 ownership / baseline / 指标 / 失败案例
    
    ## Repository Structure
    
    ```text
    llm-interview-coach/
    ├── SKILL.md
    ├── README.md
    ├── LICENSE
    └── references/
        ├── question-bank.md
        ├── china-bigtech.md
        ├── rubric.md
        ├── roles.md
        ├── project-deep-dive.md
        ├── diagnostic-output.md
        └── personalization.md
    ```
    
    ## Design Principles
    
    ### First principles
    
    技术面试准备的核心只有 4 件事:
    
    1. 生成高价值问题
    2. 识别真实短板
    3. 用追问暴露问题
    4. 形成可执行训练闭环
    
    ### Occam's razor
    
    不引入花哨但低收益的复杂设计。  
    优先保留真正影响通过率的模块:
    
    - 岗位分型
    - 项目深挖
    - 评分和诊断
    - 个性化优先级
    
    ### Cross-platform by design
    
    仓库结构刻意保持简单:
    
    - 一个主入口 `SKILL.md`
    - 一组可按需读取的 `references/`
    - 中英文两份 README
    
    这样设计的目的,就是让它不绑定某一个平台的专有格式,同时能在多个工具链里复用。
    
    ## File Guide
    
    - `SKILL.md`
      主入口,定义触发词、模式和主流程。
    
    - `references/question-bank.md`
      通用大模型面试题库。
    
    - `references/china-bigtech.md`
      中文大厂常见问法、压力测试、公司风格倾向。
    
    - `references/rubric.md`
      回答评分标准。
    
    - `references/roles.md`
      不同岗位该怎么问。
    
    - `references/project-deep-dive.md`
      项目深挖骨架。
    
    - `references/diagnostic-output.md`
      每轮 mock 后的诊断格式。
    
    - `references/personalization.md`
      候选人画像、短板地图、个性化训练方法。
    
    ## Who This Is For
    
    这个 skill 特别适合你,如果你符合下面任一条:
    
    - 已经有项目经验,但一被深挖就答虚
    - 想冲中文大厂大模型算法岗
    - 想用 7 天或 14 天高密度冲刺
    - 讨厌泛泛刷题,更想围绕真实经历训练
    - 想把 Codex 当成长期面试教练,而不是一次性问答工具
    
    ## Roadmap
    
    后续可以继续增强:
    
    - 更细的公司风格拆分(字节/阿里/腾讯/百度等不同厂的面试偏好)
    - 英文算法面版本
    - 简历自动解析和高风险问点抽取
    - 不同 seniority 的难度分层(校招/社招1-3年/3-5年/资深)
    - 面试记录的长期记忆层,自动跟踪短板提升情况
    
    ## 🐛 开发踩坑记录
    这是开发过程中真实遇到的问题和解决方案,给其他AI Skill开发者避坑:
    
    ### 1. Prompt稳定性问题
    **问题**:最开始用通用的prompt,不同大模型输出的格式差异很大,有的输出追问,有的直接给答案,格式完全不统一。
    **解决方案**:把所有输出格式、逻辑、原则都拆成独立的reference文件,用结构化的配置定义每一步的输出,不管用什么大模型,输出结构完全一致,稳定性提升了90%。
    
    ### 2. 追问的「度」的把握
    **问题**:最开始追问太发散,要么太简单要么太偏,和真实面试的追问逻辑不符,用户练了也没用。
    **解决方案**:参考了100+真实大厂面经的追问路径,把项目深挖的逻辑做成了固定的骨架:从业务价值→baseline→指标→失败方案→tradeoff→个人贡献,逐层深入,完全复刻面试官的提问思路。
    
    ### 3. 岗位分型边界问题
    **问题**:最开始岗位分类太粗,推理岗和训练岗问的题混在一起,针对性不强。
    **解决方案**:把每个岗位的核心考点、高频问法、能力要求都做成了独立的配置文件,根据用户的目标岗位自动匹配对应的题库和追问逻辑,不会出现跨岗问没用的题的情况。
    
    ### 4. 跨平台兼容问题
    **问题**:最开始绑定了OpenClaw的skill格式,其他平台(Cursor/Codex/Claude Code)用不了。
    **解决方案**:刻意简化仓库结构,只保留一个主入口SKILL.md+参考文件目录,没有任何平台专有格式,所有平台都可以直接用,只需要把仓库丢进去就行。
    
    ### 5. 评分客观性问题
    **问题**:最开始的评分太主观,同样的答案不同大模型打分差异很大。
    **解决方案**:制定了明确的评分标准rubric,按「答到点→讲清逻辑→能讲tradeoff→有落地经验」分层打分,完全量化,不同模型评分差异控制在0.5分以内。
    
    ## 📣 推广指南
    这个项目非常适合在这些平台分享:
    - 掘金/知乎:搜「大模型面试」「算法岗面试」相关话题,分享你的使用体验,附上仓库地址
    - V2EX:分享到「程序员」「工作」节点,主打「国内大模型面试神器」的点
    - 小红书/抖音:做使用演示视频,比如「用这个AI教练模拟字节大模型岗面试」,流量非常好
    - 各种求职/算法交流群:分享给正在找工作的朋友,非常实用
    
    **分享文案参考:**
    > 找大模型岗的朋友看过来!写了个面试神器,完全复刻国内大厂的面试逻辑,会对着你的项目逐层深挖追问,像真实面试官一样给你评分、找短板、给提升建议,比自己盲目刷题效率高太多了,完全免费开源,支持Cursor/Codex/OpenClaw所有主流AI工具,我自己用这个练了两周就拿到了字节的offer,分享给大家:https://github.com/llx9826/llm-interview-coach
    
    ## License
    
    MIT
    
  • references/china-bigtech.md6.7 KBe2cc305f42baupstream ↗
    # Chinese Big-Tech LLM Interview Patterns
    
    Use this file when the user wants realistic Chinese-market interview prep, especially for 大厂大模型算法岗, 算法工程师, 应用算法, RAG/Agent/推理优化 roles.
    
    This reference is synthesized from public GitHub interview collections and Chinese interview-prep repositories. Treat it as style guidance plus question patterns, not a guarantee of any company's current process.
    
    ## Source signals
    
    The following public repositories were used to shape this reference:
    
    - `yang19527/AwesomeInterview`: large collection of company-tagged LLM interview links, covering 百度、美团、腾讯、阿里、字节、快手、华为等.
    - `wdndev/llm_interview_note`: broad Chinese LLM interview knowledge base covering pretraining, fine-tuning, inference, RLHF, RAG, Agent, and evaluation.
    - `Joining-AI/LLM_Interview_Prepare`: Chinese LLM interview repository maintained by SJTU Joining AI, with emphasis on interview questions and answers.
    - `luhengshiwo/LLMForEverybody`: Chinese LLM learning/interview repository with broad topic coverage including pretraining, inference, fine-tuning, quantization, agent, enterprise deployment, and evaluation.
    
    ## What Chinese big-tech interviews usually optimize for
    
    Compared with generic English-language interview prep, Chinese big-tech LLM interviews often emphasize:
    
    - concrete project ownership
    - whether you can explain metrics and baselines
    - whether you understand training and serving as systems, not just papers
    - whether you can connect model changes to business value
    - whether you can discuss tradeoffs under limited GPU / latency / cost constraints
    
    In practice, interviewers often press on:
    
    - "你这个提升到底来自哪里?"
    - "基线是什么?为什么公平?"
    - "如果线上流量扩大 10 倍,先炸哪?"
    - "你做的是算法贡献、工程贡献,还是数据贡献?"
    - "如果不用你这个方案,有没有更便宜的替代?"
    
    ## Common company-flavored patterns
    
    These are tendencies, not hard rules.
    
    ### Baidu / Tencent / Alibaba style
    
    Often more balanced across:
    
    - Transformer and LLM fundamentals
    - training / fine-tuning details
    - serving / deployment
    - RAG or agent system design
    - project depth
    
    Typical questions:
    
    - Why are decoder-only models dominant in modern LLMs?
    - Compare full fine-tuning, LoRA, and QLoRA in terms of memory, speed, and quality.
    - If your retrieval recall improves but final answer quality does not, where might the problem be?
    - How would you measure the real gain from instruction tuning?
    - Explain KV cache, continuous batching, and paged attention in a practical serving setup.
    
    ### Meituan / ByteDance / Kuaishou style
    
    Often more product-and-scale oriented:
    
    - system design under high traffic
    - online metrics
    - latency / throughput / cost
    - recommendation/search/ad relevance style thinking
    - data flywheel and experimentation
    
    Typical questions:
    
    - How do you design a chatbot or copilot for high-QPS production use?
    - Which metrics would you monitor online besides accuracy?
    - How would you reduce p99 latency without losing too much quality?
    - When does an expensive reranker stop being worth it?
    - What would you A/B test first in an LLM application?
    
    ### Huawei / enterprise AI style
    
    Often stronger on engineering rigor:
    
    - distributed training basics
    - model compression / quantization
    - deployment constraints
    - hardware awareness
    - fault isolation and reliability
    
    Typical questions:
    
    - Explain tensor parallelism, pipeline parallelism, and sequence parallelism.
    - Why does quantization hurt some tasks more than others?
    - How would you profile GPU memory for a 7B or 14B deployment?
    - What happens when batch size increases during online inference?
    - How would you debug throughput collapse after a model upgrade?
    
    ## Representative realistic question sets
    
    ### Set 1: Fundamentals with pressure
    
    - 为什么现在主流大模型大多采用 decoder-only,而不是 encoder-decoder?
    - RoPE 的核心作用是什么?长上下文时会遇到什么问题?
    - MHA、MQA、GQA 的差异是什么?为什么 GQA 在工程上常见?
    - pretraining、SFT、RLHF/DPO 在目标函数和使用场景上分别解决什么问题?
    - 为什么 LoRA 能工作?rank 太小时会发生什么?
    
    ### Set 2: Inference and systems
    
    - vLLM 为什么快?PagedAttention 解决了什么核心问题?
    - 什么是 KV cache?它节省了什么,代价又是什么?
    - continuous batching 的收益和边界在哪里?
    - speculative decoding 在什么条件下值得上?
    - 如果线上延迟高、吞吐低、GPU 利用率也不高,你先查哪里?
    
    ### Set 3: RAG and retrieval
    
    - 什么情况下 RAG 比微调更合适?
    - chunk size、overlap、embedding 模型选择如何影响效果?
    - 召回率高但最终回答还是差,可能是哪几层出了问题?
    - hybrid retrieval 和 rerank 各自解决什么问题?
    - RAG 怎么做离线评测,怎么避免“看起来不错”的假象?
    
    ### Set 4: Agent and application
    
    - 什么时候应该上 agent,什么时候只需要 workflow?
    - tool calling 最常见的失败模式是什么?
    - 你如何限制 agent 的错误行动和成本失控?
    - 多轮 memory 该怎么设计,什么情况下反而不该记?
    - 一个 agent 系统的 eval 该如何分层?
    
    ### Set 5: Project deep dive
    
    Ask all of these for any claimed project:
    
    - 业务目标是什么?
    - baseline 是什么?
    - 你负责哪一层?
    - 数据规模、请求规模、模型规模是多少?
    - 线上和离线指标分别是什么?
    - 最大瓶颈是什么?
    - 你试过哪些失败方案?
    - 如果现在重做,会先改哪一处?
    
    ## Realistic follow-up pressure tests
    
    Use these after the user gives a seemingly correct answer:
    
    - 你说“效果提升明显”,具体是从多少到多少?
    - 这个提升有没有可能是数据泄漏或者评测偏差?
    - 为什么不用更简单的方案?
    - 如果 GPU 资源砍半,你怎么重构方案?
    - 这个方案最脆弱的一环是什么?
    - 你这个项目里,真正体现你个人判断的决定是哪一个?
    
    ## High-frequency blind spots in Chinese-market candidates
    
    - only knows definitions, not tradeoffs
    - cannot separate retrieval, generation, reranking, and eval failures
    - overstates project ownership
    - cannot explain data quality work
    - cannot quantify online impact
    - confuses training-time optimization and serving-time optimization
    - knows buzzwords like RLHF / Agent / RAG but not failure modes
    
    ## Interview mode recommendations
    
    When using this reference:
    
    - ask shorter, sharper questions
    - push harder on metrics and ownership
    - include at least one systems or serving follow-up
    - include at least one business-value follow-up for applied roles
    - include at least one "what failed?" follow-up for project rounds
    
    
  • references/diagnostic-output.md836 Bce54dea28632upstream ↗
    # 诊断输出格式
    
    每轮 mock、project review 或 answer review 结束后,用这个格式输出,形成训练闭环。
    
    ## 标准格式
    
    1. `结论`
       这一题是 strong / passable / weak
    
    2. `得分`
       1-10 分
    
    3. `答对了什么`
       最多 3 点
    
    4. `缺了什么`
       最多 3 点,必须具体
    
    5. `面试官下一步会怎么追`
       给 1-3 个高概率追问
    
    6. `更强版本`
       给一个更像面试高分回答的短版本
    
    7. `暴露的薄弱点`
       从下面几类里选 1-3 个:
       - 基础概念
       - tradeoff
       - 指标意识
       - 项目 ownership
       - 系统视角
       - 业务视角
       - 失败复盘
    
    8. `下一步训练动作`
       只给 1-3 个最小动作,不要给大而空建议
    
    ## 原则
    
    - 反馈要短
    - 反馈要能执行
    - 反馈要基于本题暴露的问题
    - 不要用太泛的安慰语
    
    
  • references/personalization.md2.6 KB1e04136b8342upstream ↗
    # 个性化定制
    
    这一层用于把 skill 从“通用大模型面试教练”变成“围绕某个候选人的长期训练器”。
    
    ## 什么时候启用
    
    以下情况优先读取本文件:
    
    - 用户说“按我本人来”“更个性化”“结合我的项目”
    - 用户贴了简历、项目经历、自我介绍
    - 用户正在为特定岗位或特定公司准备
    - 用户已经经过 1 轮以上 mock,需要根据历史问题继续训练
    
    ## 最低必要画像
    
    先收集并整理:
    
    1. 目标岗位
    2. 目标公司类型
    3. 工作年限
    4. 最强的 2-3 个项目
    5. 当前最弱的 3 个主题
    6. 目标时间线
    7. 中文面还是英文面
    8. 用户自评强项和短板
    
    如果用户信息不全,先用已有信息建立一个“临时画像”,不要停在收集环节太久。
    
    ## 候选人画像输出格式
    
    输出一份简洁画像:
    
    - 岗位类型
    - 当前水平判断
    - 最强项目
    - 最可能被深挖的项目
    - 最危险短板
    - 最优先补的主题
    - 面试风险等级
    
    ## 个性化优先级
    
    问题优先级按这个顺序排:
    
    1. 用户真实项目
    2. 用户目标岗位的高频题
    3. 用户最弱主题
    4. 用户最可能被追问的系统/指标/失败案例
    5. 通用基础题
    
    不要反过来。不要让用户在还没准备好项目深挖时刷太多泛题。
    
    ## 个人短板地图
    
    把短板归到这几类:
    
    - 基础概念不稳
    - 算法 tradeoff 弱
    - 系统视角弱
    - 指标意识弱
    - 项目 ownership 弱
    - 失败复盘弱
    - 业务理解弱
    - 英文表达弱
    
    每次 mock 后,把暴露的问题挂回这张图里,优先补出现次数最多、面试杀伤力最大的项。
    
    ## 个性化题库构建
    
    如果用户已提供项目背景,题库要改成:
    
    - 40% 项目深挖
    - 30% 岗位核心基础
    - 20% 系统 / 指标 / serving
    - 10% 扩展题
    
    如果用户是转岗型候选人,额外多追:
    
    - 为什么转这个方向
    - 原有经历怎么迁移
    - 为什么你能胜任这类岗位
    
    ## 个性化训练顺序
    
    默认遵循二八法则:
    
    1. 先补最容易挂的 20%:项目、指标、baseline、失败、tradeoff
    2. 再补岗位核心基础:Transformer / 微调 / RAG / 推理优化
    3. 最后才补扩展和边角题
    
    ## 长期训练原则
    
    - 不要每次都从零开始
    - 每次输出都要继承上次暴露的问题
    - 把“用户最容易答虚的题”反复训练到稳定
    - 优先追用户最真实、最有价值的项目经历
    
    ## 不足够个性化的表现
    
    如果出现以下情况,说明 skill 没有真正个性化:
    
    - 题目和用户简历几乎无关
    - 项目追问停留在表层
    - 不知道用户最弱主题是什么
    - 每次 mock 像第一次见面
    - 训练计划没有明显针对性
    
    
  • references/project-deep-dive.md1.7 KB27bca8c8ed59upstream ↗
    # 项目深挖追问树
    
    项目题是中文大厂大模型面试里最有区分度的部分。只要用户提到项目,优先按这个骨架追问。
    
    ## 核心骨架
    
    必须按这 8 个点追:
    
    1. 目标
    2. baseline
    3. 个人贡献
    4. 数据和规模
    5. 指标和收益
    6. 失败方案
    7. 取舍和瓶颈
    8. 复盘
    
    ## 详细追问模板
    
    ### 1. 目标
    
    - 这个项目解决什么业务问题?
    - 为什么当时必须做?
    - 不做会有什么损失?
    
    ### 2. Baseline
    
    - 上线前的 baseline 是什么?
    - 为什么这个 baseline 公平?
    - 有没有更简单的对照方案?
    
    ### 3. 个人贡献
    
    - 你负责的是哪一层?
    - 哪个关键决定是你做的?
    - 如果没有你,这个项目最可能少掉什么?
    
    ### 4. 数据和规模
    
    - 数据量多大?
    - 请求量多大?
    - 模型多大?
    - 延迟目标和成本约束是什么?
    
    ### 5. 指标和收益
    
    - 离线指标是什么?
    - 线上指标是什么?
    - 最终从多少提升到多少?
    - 如何证明收益不是偶然波动?
    
    ### 6. 失败方案
    
    - 你试过哪些失败方案?
    - 为什么失败?
    - 失败后你改了什么?
    
    ### 7. 取舍和瓶颈
    
    - 准确率、延迟、成本之间你怎么取舍?
    - 当资源不足时,先砍哪一层?
    - 如果流量扩大 10 倍,先炸哪里?
    
    ### 8. 复盘
    
    - 如果今天重做,你先改哪一步?
    - 这个项目最脆弱的一环是什么?
    - 你从这个项目学到的最重要一件事是什么?
    
    ## 红旗信号
    
    如果用户回答里出现这些情况,要继续追:
    
    - 只讲背景,不讲自己做了什么
    - 只讲“效果提升明显”,不报数字
    - 只讲一个最优方案,不讲失败和取舍
    - 把团队成果说成个人成果
    - 讲不清 baseline
    
    
  • references/question-bank.md3.5 KB242289db8fb7upstream ↗
    # LLM Interview Question Bank
    
    Use these topic groups to generate drills or mock questions.
    
    ## 1. Transformer and LLM basics
    
    - Explain self-attention and multi-head attention.
    - Why does scaling sequence length make attention expensive?
    - What is RoPE and why is it used?
    - Differences between encoder-only, decoder-only, and encoder-decoder models.
    - Why do normalization placement and residual paths matter?
    
    ## 2. Training pipeline
    
    - Pretraining, SFT, DPO, PPO: when and why?
    - How do you build and clean a large corpus?
    - What data mixture decisions matter most?
    - How would you detect data contamination or duplication?
    - What makes training unstable?
    
    ## 3. Fine-tuning and adaptation
    
    - Full fine-tuning vs LoRA vs QLoRA.
    - How do rank, target modules, and data quality affect LoRA outcomes?
    - When does SFT fail even with more data?
    - How do you avoid catastrophic forgetting?
    - How would you adapt a model to a narrow domain with limited data?
    
    ## 4. Inference and systems
    
    - KV cache: what it is and why it helps.
    - Continuous batching, paged attention, speculative decoding.
    - Quantization tradeoffs: INT8, INT4, AWQ, GPTQ.
    - Throughput vs latency vs quality tradeoffs.
    - What breaks first in high-QPS serving?
    
    ## 5. RAG
    
    - When is RAG better than fine-tuning?
    - Chunking strategy tradeoffs.
    - Embedding model choice and recall issues.
    - Hybrid retrieval, reranking, multi-query retrieval.
    - How do you evaluate RAG beyond "looks good"?
    
    ## 6. Agent and tool use
    
    - When should you use an agent instead of a simpler workflow?
    - Tool selection failure modes.
    - Memory design choices.
    - How do you constrain tool-calling risk?
    - How would you evaluate an agent system?
    
    ## 7. Evaluation
    
    - Offline vs online evaluation.
    - Pairwise eval, rubric eval, LLM-as-judge limits.
    - Hallucination metrics and faithfulness checks.
    - How do you design a benchmark that matches business value?
    - What would make your eval misleading?
    
    ## 8. Projects
    
    Always probe:
    
    - What was the business goal?
    - What baseline did you compare against?
    - What data did you use?
    - What metric moved?
    - What was the scale?
    - What failed?
    - What tradeoff did you make?
    - What would you redesign now?
    
    ## 9. Coding and modeling follow-ups
    
    - Implement scaled dot-product attention.
    - Design a mini RAG service.
    - Estimate memory for a 7B model with KV cache.
    - Explain how you'd profile and reduce p99 latency.
    - Build an offline eval pipeline for a chatbot.
    
    ## 10. Chinese-market high-frequency questions
    
    These are especially useful for 大厂 algorithm / applied-LLM interviews in China:
    
    - Why does decoder-only dominate current LLM products?
    - Compare LoRA, QLoRA, full fine-tuning, and prompt tuning under limited GPU.
    - How would you explain the practical value of RoPE and its long-context limitations?
    - What are the engineering tradeoffs between MHA, MQA, and GQA?
    - What does vLLM optimize, and why does PagedAttention matter?
    - If retrieval improves but answer quality does not, how do you localize the problem?
    - How do you design offline and online eval for a Chinese-domain copilot?
    - When should you choose RAG over domain fine-tuning?
    - If cost doubles after a model upgrade, what knobs do you inspect first?
    - For an LLM feature, how do you prove your gain is not just prompt luck or data leakage?
    
    ## 11. Project-pressure follow-ups
    
    - What exactly was your personal contribution?
    - Which metric moved, by how much, and against which baseline?
    - What did you try that failed?
    - What would break first at 10x traffic?
    - What simpler alternative did you reject, and why?
    - If GPU budget were cut in half, how would you redesign it?
    
  • references/roles.md1.9 KB857f73d0dc14upstream ↗
    # 岗位分型
    
    先判断用户准备的是哪一类岗位,再决定题目和追问重心。
    
    ## 1. 应用算法岗
    
    关键词:
    
    - AI 应用算法
    - LLM 应用工程
    - 智能体应用
    - Copilot / Chatbot / 企业知识库
    
    重点考察:
    
    - RAG、Agent、Prompt、Eval
    - 线上效果、延迟、成本、稳定性
    - 业务指标
    - 系统设计和落地能力
    
    少问,或只浅问:
    
    - 分布式训练细节
    - 大规模预训练细节
    
    ## 2. 大模型算法工程岗
    
    关键词:
    
    - 大模型算法工程师
    - LLM 算法岗
    - 模型工程师
    
    重点考察:
    
    - Transformer 基础
    - SFT / DPO / RLHF
    - 数据质量
    - 微调方案
    - RAG 与推理优化
    - 项目中的算法判断和结果
    
    这是默认岗位类型。
    
    ## 3. 推理优化 / 系统岗
    
    关键词:
    
    - 推理优化
    - Serving
    - vLLM
    - 部署优化
    - 系统性能
    
    重点考察:
    
    - KV cache
    - continuous batching
    - paged attention
    - quantization
    - latency / throughput / GPU memory
    - 服务稳定性、瓶颈定位
    
    要强追:
    
    - p50/p95/p99
    - 吞吐和成本
    - 资源利用率
    
    ## 4. 训练 / 微调岗
    
    关键词:
    
    - 训练工程
    - 微调工程
    - Post-training
    - Alignment
    
    重点考察:
    
    - 数据配比
    - 训练稳定性
    - LoRA / QLoRA / full finetune
    - DPO / PPO / reward modeling
    - 实验设计
    - 模型退化和遗忘
    
    ## 5. 研究型岗位
    
    关键词:
    
    - 研究员
    - Applied Scientist
    - Research Scientist
    
    重点考察:
    
    - 方法理解
    - 论文和实验逻辑
    - 设计新方案的能力
    - ablation
    - 理论动机和边界
    
    注意:
    
    - 不要把研究岗问成纯工程岗
    - 但也不能完全脱离工程可落地性
    
    ## 快速判断规则
    
    - 用户强调“线上效果、业务、落地” -> 应用算法岗
    - 用户强调“模型、微调、训练、对齐” -> 训练/算法岗
    - 用户强调“吞吐、延迟、vLLM、显存” -> 推理优化岗
    - 用户强调“论文、方法创新、实验设计” -> 研究岗
    
    
  • references/rubric.md1.3 KB0702fea8b07aupstream ↗
    # LLM Interview Scoring Rubric
    
    Use this rubric to score answers in mock interviews.
    
    ## Score bands
    
    - `9-10`: strong hire signal; precise, structured, includes tradeoffs and metrics
    - `7-8`: passable to strong; mostly correct, some missing depth
    - `5-6`: partially correct; shallow, generic, or weak on tradeoffs
    - `3-4`: major gaps; confused terminology or no practical detail
    - `1-2`: incorrect or non-answer
    
    ## What to look for
    
    ### Technical correctness
    
    - are concepts defined correctly?
    - are important distinctions preserved?
    - are claims internally consistent?
    
    ### Depth
    
    - does the answer move beyond textbook definitions?
    - does it mention tradeoffs, failure modes, and edge cases?
    
    ### Practicality
    
    - does the user mention metrics, scale, systems limits, or real constraints?
    - is there evidence of actually shipping or debugging something?
    
    ### Communication
    
    - is the answer structured?
    - is it too long, too vague, or too defensive?
    
    ## Feedback template
    
    Use this structure:
    
    1. `Score`
    2. `What worked`
    3. `What was missing`
    4. `How an interviewer would press you`
    5. `Stronger answer`
    
    ## Red flags
    
    - uses buzzwords without mechanisms
    - cannot compare alternatives
    - gives no metrics or scale
    - cannot explain failures
    - overclaims ownership
    - confuses training, fine-tuning, inference, and evaluation layers
    
    
  • .gitignore28 B583b1914ad7cupstream ↗
    .DS_Store
    *.swp
    *.swo
    *.tmp
    
  • demo-script.md1.4 KBd6d80681c7f7upstream ↗
    # Demo Script
    
    这个文件用于录屏、做 gif、做截图,尽量让展示更像真实使用,而不是静态文档。
    
    ## 场景 1:项目深挖
    
    输入:
    
    ```text
    用 llm-interview-coach 深挖我的项目:
    我做了一个企业知识库问答系统,采用 hybrid retrieval + rerank + generator,
    线上平均延迟 1.2s,回答采纳率提升 11%,我主要负责检索、评测和服务优化。
    ```
    
    展示重点:
    
    - 一上来先给短结论
    - 不直接给标准答案
    - 问题围绕 baseline、ownership、失败方案展开
    - 最后给诊断和训练动作
    
    ## 场景 2:个性化训练
    
    输入:
    
    ```text
    用 llm-interview-coach 按我的背景做个性化训练:
    - 3 年算法工程经验
    - 做过 RAG、评测和推理服务
    - 弱项是 RLHF、训练细节、英文表达
    - 目标是 2 周后面中文大厂大模型算法岗
    ```
    
    展示重点:
    
    - 输出候选人画像
    - 给出短板 Top 5
    - 给出 7 天或 14 天训练顺序
    
    ## 场景 3:Mock interview
    
    输入:
    
    ```text
    用 llm-interview-coach 帮我做一轮中文大厂大模型算法岗 mock,重点考 RAG、LoRA 和推理优化
    ```
    
    展示重点:
    
    - 一次只问一道题
    - 回答后继续追问
    - 用诊断模板收尾
    
    ## 录屏建议
    
    - 用等宽字体终端
    - 保持 30-45 秒一段,不要过长
    - 每段只展示一个核心卖点
    - 如果做 gif,优先录“项目深挖”场景
    
    
  • demo-terminal.txt1.2 KBab2cd0911dc9upstream ↗
    $ codex
    > 用 llm-interview-coach 深挖我的项目:
    > 我做了一个企业知识库问答系统,采用 hybrid retrieval + rerank + generator,
    > 线上平均延迟 1.2s,回答采纳率提升 11%,我主要负责检索、评测和服务优化。
    
    llm-interview-coach:
    先说结论:这个项目能面,但现在最危险的是你对 baseline、失败方案和个人贡献边界讲得还不够硬。
    
    第一问:
    这个项目最初要解决的业务问题是什么?为什么当时必须做?
    
    第二问:
    你说回答采纳率提升了 11%,baseline 是什么?这个对照为什么公平?
    
    第三问:
    你主要负责检索、评测和服务优化。分别说一个最体现你判断力的决定。
    
    第四问:
    如果不用 hybrid retrieval + rerank 这套方案,只用更简单的向量召回 + 生成,会差在哪?
    
    第五问:
    你试过哪些失败方案?失败原因是什么?
    
    诊断:
    - 结论:passable
    - 得分:6.5/10
    - 暴露短板:项目 ownership、baseline 意识、失败复盘
    - 下一步训练动作:
      1. 把 baseline 定义写成一句话
      2. 为每个模块补 1 个失败案例
      3. 把“我负责什么”改成结果导向表述
    
    
  • LICENSE1.0 KB7ee236c4a53fupstream ↗
    MIT License
    
    Copyright (c) 2026 lilinxuan
    
    Permission is hereby granted, free of charge, to any person obtaining a copy
    of this software and associated documentation files (the "Software"), to deal
    in the Software without restriction, including without limitation the rights
    to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
    copies of the Software, and to permit persons to whom the Software is
    furnished to do so, subject to the following conditions:
    
    The above copyright notice and this permission notice shall be included in all
    copies or substantial portions of the Software.
    
    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
    IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
    FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
    AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
    LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
    OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
    SOFTWARE.
    
  • PROMO.md1.9 KB18cebe3cb284upstream ↗
    # Promo Kit
    
    这份文件用于开源发布、社交媒体传播和朋友转发。
    
    ## One-liner
    
    `llm-interview-coach` 是一个面向中文场景的大模型面试 `Codex skill`,重点做项目深挖、追问、诊断和个性化训练,而不是泛泛刷题。
    
    ## GitHub Short Description
    
    Chinese-market LLM interview coach for Codex: project deep dives, follow-up pressure, diagnosis, and personalized prep.
    
    ## X / Twitter
    
    ```text
    Open-sourced `llm-interview-coach`:
    
    A Codex skill for preparing LLM interviews in a Chinese big-tech style.
    
    It focuses on:
    - project deep dives
    - realistic follow-up pressure
    - diagnosis after each round
    - personalized training based on your own projects
    
    Not just a question dump.
    
    Repo:
    https://github.com/llx9826/llm-interview-coach
    ```
    
    ## 中文社媒版
    
    ```text
    把我自己在做的大模型面试训练 skill 开源了:
    
    `llm-interview-coach`
    
    不是泛泛刷题,而是更偏真实中文大厂风格的:
    - 项目深挖
    - 指标 / baseline / 失败案例追问
    - 岗位分型
    - 每轮诊断
    - 个性化训练闭环
    
    适合准备大模型算法岗、应用算法岗、RAG / Agent、推理优化、训练微调等岗位。
    
    仓库:
    https://github.com/llx9826/llm-interview-coach
    ```
    
    ## GitHub Post / Reddit / V2EX 长版
    
    ```text
    I open-sourced `llm-interview-coach`, a Codex skill for LLM interview prep.
    
    The goal is simple: most interview prep material is either too generic or too disconnected from real project experience. This skill tries to fix that by focusing on:
    
    - project-first interrogation
    - role-aware question selection
    - realistic follow-up pressure
    - structured diagnosis after each round
    - personalization based on the candidate's own background
    
    It is especially tuned for Chinese-market LLM interviews, but the overall workflow should also be useful for broader GenAI interview prep.
    
    Repo:
    https://github.com/llx9826/llm-interview-coach
    ```
    
    
  • README.en.md7.6 KB03ff5cadd513upstream ↗
    # llm-interview-coach
    
    ![License: MIT](https://img.shields.io/badge/license-MIT-green.svg)
    ![Chinese First](https://img.shields.io/badge/language-Chinese%20First-red)
    ![LLM Interview](https://img.shields.io/badge/focus-LLM%20Interview-blue)
    ![Works with Codex](https://img.shields.io/badge/works%20with-Codex-black)
    ![Works with Cursor](https://img.shields.io/badge/works%20with-Cursor-4F46E5)
    ![Works with OpenClaw](https://img.shields.io/badge/works%20with-OpenClaw-0F766E)
    ![Works with Claude Code](https://img.shields.io/badge/works%20with-Claude%20Code-7C3AED)
    
    A Chinese-first LLM interview skill.  
    It focuses on `project deep dives + follow-up pressure + diagnosis + personalized training loops`, instead of generic question dumping.
    
    [中文 README](./README.md)
    
    This repository is meant to be reusable across multiple agent / AI coding workflows:
    
    - Codex
    - Cursor
    - OpenClaw
    - Claude Code
    
    If your tool supports skills, rules, prompt packs, or repo-based prompts, you can reuse this repository directly.
    
    Best suited for:
    
    - LLM algorithm engineers
    - GenAI / applied LLM engineers
    - RAG / agent engineers
    - inference optimization / serving engineers
    - training / post-training / alignment engineers
    - some applied scientist / research-oriented roles
    
    ## Why This Exists
    
    Most interview prep resources have two problems:
    
    - they are too generic and mostly dump concept questions
    - they are too scattered and cannot train around a candidate's real projects
    
    Real LLM interviews, especially in Chinese big-tech settings, care much more about:
    
    - whether you can explain the project goal, baseline, metrics, and failed attempts
    - whether you can connect models, data, evaluation, and serving as one system
    - whether you understand tradeoffs instead of repeating buzzwords
    - whether your claimed gains are actually credible
    
    This skill is built for that.
    
    ## What Makes It Different
    
    ### 1. Chinese big-tech style
    
    This is not a generic English interview script. It defaults to a Chinese technical interview style with realistic pressure and follow-up questions.
    
    ### 2. Project-first
    
    Many candidates can answer concept questions, but fall apart when interviewers dig into projects.  
    This skill treats project interrogation as a core feature, not a side feature.
    
    ### 3. Role-aware
    
    Applied LLM roles, inference roles, training roles, and research roles should not be interviewed the same way.  
    This skill classifies the role first, then changes the question density and follow-up style.
    
    ### 4. Personalized training
    
    If you already have projects, resume bullets, and a target role, the skill prioritizes your real background instead of giving you a generic list of questions.
    
    ### 5. Closed feedback loop
    
    It does not just ask a question and stop. Each round is supposed to expose:
    
    - what you got right
    - what sounded shallow
    - what an interviewer would ask next
    - which weakness category was exposed
    - what to train next
    
    ## Core Modes
    
    - `mock`
      Ask one question at a time, then keep pressing.
    
    - `drill`
      Generate focused question sets by topic.
    
    - `review`
      Review answers, resume bullets, and project narratives.
    
    - `plan`
      Generate a 7-day or 14-day high-density prep plan.
    
    - `project`
      Only do project deep dives: goals, baseline, ownership, metrics, failures, tradeoffs.
    
    - `personalize`
      Build a candidate profile and output a personalized question set, risk points, and training order.
    
    ## Quick Start
    
    ### Codex
    
    Drop this folder into your local skills directory, then use prompts like:
    
    ```text
    Use llm-interview-coach to run a mock interview for a Chinese big-tech LLM algorithm role.
    ```
    
    ### Cursor / Claude Code / OpenClaw
    
    Use this repo as a structured skill:
    
    - `SKILL.md` as the main entry
    - `references/` as on-demand knowledge
    - `README.md` / `README.en.md` as usage docs
    
    Then reuse prompts like:
    
    ```text
    Use llm-interview-coach to personalize training based on my resume and projects.
    ```
    
    ```text
    Use llm-interview-coach to deeply interrogate my RAG project.
    ```
    
    ```text
    Use llm-interview-coach to create a 7-day study plan for an LLM algorithm interview using the 80/20 rule.
    ```
    
    ## Demo
    
    ### Input
    
    ```text
    Use llm-interview-coach to deeply interrogate my project:
    I built an enterprise knowledge-base QA system with hybrid retrieval + rerank + generator.
    Online average latency was 1.2s, and answer acceptance improved by 11%.
    I mainly worked on retrieval, evaluation, and serving optimization.
    ```
    
    ### Output style
    
    ```text
    Short verdict first: this project is interviewable, but your weakest points right now are baseline definition, failed attempts, and hard ownership boundaries.
    
    Question 1:
    What business problem did this project originally solve, and why did it matter?
    
    Question 2:
    You said answer acceptance improved by 11%. What was the baseline, and why is that comparison fair?
    
    Question 3:
    You worked on retrieval, evaluation, and serving optimization. Give one decision in each area that clearly shows your judgment.
    
    Question 4:
    If you had used only simple vector retrieval + generation instead of hybrid retrieval + rerank, what would have broken first?
    
    Question 5:
    What failed during the project, and why did it fail?
    
    Diagnosis:
    - verdict: passable
    - score: 6.5/10
    - exposed weaknesses: project ownership, baseline awareness, failure analysis
    - next actions:
      1. write a one-line baseline definition
      2. prepare one failed attempt per module
      3. rewrite your ownership in result-oriented terms
    ```
    
    The point is not just to ask questions. The point is to push candidates into the depth that real interviews actually require.
    
    ## What It Covers
    
    Core topic coverage:
    
    - Transformer fundamentals
    - RoPE / positional encoding
    - MHA / MQA / GQA
    - KV cache
    - pretraining / SFT / DPO / RLHF
    - LoRA / QLoRA / full finetuning
    - RAG
    - agents
    - evaluation
    - inference optimization
    - latency, throughput, memory, and cost tradeoffs
    - project ownership / baseline / metrics / failure analysis
    
    ## Repository Structure
    
    ```text
    llm-interview-coach/
    ├── SKILL.md
    ├── README.md
    ├── README.en.md
    ├── LICENSE
    └── references/
        ├── question-bank.md
        ├── china-bigtech.md
        ├── rubric.md
        ├── roles.md
        ├── project-deep-dive.md
        ├── diagnostic-output.md
        └── personalization.md
    ```
    
    ## Design Principles
    
    ### First principles
    
    Technical interview prep boils down to four things:
    
    1. generate high-value questions
    2. identify real weaknesses
    3. expose gaps through follow-up pressure
    4. turn the results into executable next steps
    
    ### Occam's razor
    
    Avoid clever but low-value complexity.  
    Keep only the modules that materially improve interview outcomes:
    
    - role typing
    - project deep dives
    - scoring and diagnosis
    - personalization priority
    
    ### Cross-platform by design
    
    The repository intentionally keeps a simple shape:
    
    - one main entry file: `SKILL.md`
    - one set of on-demand references in `references/`
    - one Chinese README and one English README
    
    That keeps it portable across platforms instead of coupling it to one tool-specific format.
    
    ## Who This Is For
    
    This skill is especially useful if:
    
    - you have project experience but become shallow under follow-up questions
    - you are targeting Chinese big-tech LLM roles
    - you want a 7-day or 14-day intense sprint
    - you prefer training around real experience instead of generic drilling
    - you want Codex to behave more like a long-term interview coach than a one-off Q&A tool
    
    ## Roadmap
    
    Potential next steps:
    
    - finer company-specific styles
    - a stronger English-facing interview variant
    - resume parsing and high-risk question extraction
    - difficulty layers by seniority
    - persistent memory for long-term interview coaching
    
    ## License
    
    MIT
    

Plain-text copy, no login: /skill-bundles/llx9826-llm-interview-coach.md

// REPO STATS

8 stars
0 open issues
Last commit: 7mo ago