⏳ This skill is pending AI review.
Scores will appear once the review pipeline completes.
job-research
Research public company, product-line, business, market, and competitor context for one saved job when the user explicitly requests it. Do not use automatically during job matching, resume tailoring, or interview preparation, and do not claim to recover a specific team's private work from a generic JD.
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.
// RATINGS
Not yet listed on ClawHub or SkillsMP
// README
MyCareer
可用,持续迭代中
跑在你自己电脑上的 AI 求职助手
找岗位 · 看匹配 · 改简历 · 练面试 · 管投递

求职时,JD 在浏览器里、简历在文件夹里、投递进度在脑子里、面试安排在邮箱里。MyCareer 把它们放进同一个本地工作区:你用对话把事情交给 Agent,它去查岗位、比简历、改草稿、陪你练面试、记进度。
- 数据本地保存:岗位、简历、投递和对话都存在你电脑上的 SQLite 里,没有中心服务器。
- 匹配证据可核对:每条"符合"都引用你简历里的原文;没有证据的地方只会标成缺口,不会替你编经历。
- 外部操作由你确认:写日历、改记忆这类动作,都要你点确认才会执行。
功能
🔍 找到值得投的岗位
- 岗位收藏:装上 Chrome 扩展后,在 BOSS 直聘岗位详情页点一下,就把完整 JD 存进岗位库。只有你点击才会保存,扩展不会自动搜索或投递。
- 岗位分析:把 JD 要求分成四档:硬门槛、核心要求、重要加分、普通加分,每一条都附上对应的 JD 原文。分错了可以改,系统会同时保留原判断和你的修正。
- 岗位比较:结合你确认过的求职偏好,比较几个岗位哪个更值得继续看。覆盖率的算法是公开的:完全符合算 1、部分符合算 0.5、缺失算 0,看不出来的不计入。
- 公司调研:针对岗位库里的岗位检索公开资料,每条结论都标明来源,并区分"事实""推断"和"未知"。
🧩 看清匹配,找到差距
- 简历管理:导入 PDF、TXT 或 Markdown。每次导入都是一个不可改的版本,可以随时打开原文件。简历按目标岗位分组,也能移到别的岗位下;重复上传同一个文件不会存两份。
- 简历点评:不需要 JD,按简历所在的目标岗位,从 HR 初筛、面试官和简历编辑三个角度指出写得弱的地方和改法,并列出面试官可能追问的问题。扫描件或特殊字体的 PDF 读不出文字时,会自动交给模型识别。
- 逐条匹配:拿指定的简历版本对照岗位的每条要求,结论只有四种:匹配、部分匹配、未体现、不明确。整体结论分为强、中、弱和证据不足,这不是录用概率。
- 补强建议:针对缺口给出下一步:澄清问题、补材料、做项目或学技能。
✍️ 把简历和面试准备好
- 简历定制:针对目标岗位生成修改草稿,每条修改都必须引用原简历的内容。草稿先给你看,后台同时自动审核有没有夸大;审核通过后逐条接受或拒绝,再下载定稿。
- 面试准备:根据投递和面试轮次,整理相关经历和项目素材。
- 模拟面试:
- 可以针对某次投递练,也可以自由练习。自由练习开始前会先问你用哪份简历(也可以现场上传,或者不用简历),可以指定岗位库里的岗位,也可以只说一个公司。
- 开始时先按岗位出好整场的题,逐题提问,根据你的回答决定要不要追问。默认 10 道主问题,每题最多追问 2 次。
- 常见大厂(字节、阿里、腾讯等)会参考它们的面试侧重。如果你做过这家公司的调研,也会结合调研里的业务背景出题。
- 整场结束后逐题评估:从准确性、相关性、具体性、结构、推理和表达六个方面,每项 1–5 分,再生成复盘报告,并指出前后回答里互相矛盾的地方。
- 中途离开,可以从面试中心回到原对话接着答;练完或取消的练习可以"再来一次"。
📮 把求职进度管起来
- 投递追踪:在外部平台投递后,在这里记一笔:选岗位、选当时用的简历版本、填投递时间。之后跟踪状态、事件和面试轮次。
- 每日概览:Dashboard 和日报汇总进展和下一步要做的事。
- 招聘邮件:连接 Gmail(只读)或 QQ 邮箱,识别招聘相关消息,由你确认后再更新进度。
- 面试日历:在面试中心的月历里查看安排;连接 Google 日历后,确认一次写入一次。
💬 对话与记忆
- 对话式操作:直接说出你想做什么。需要你选择或补充信息时,会弹出选项卡片;想自己填写,选"其他"就能在同一行输入。
- 长期记忆:只保存你确认过的经历和偏好,可以修改、可以删除。Agent 自己的推断会单独存放,不会直接影响推荐或匹配。
界面预览
快速开始
需要: Python 3.11+ · uv · Node.js
^20.19.0或>=22.12.0· 一个 HTTPS 的 OpenAI 兼容模型接口 收藏 BOSS 岗位还需要 Chrome、Edge、Brave、Arc 等 Chromium 内核浏览器。 Windows 上请把仓库放在路径较短的目录(如C:\code\MyCareer),或开启系统长路径支持;部分依赖的文件路径很长,目录太深会超过 260 字符上限而导入失败。
1. 安装依赖
uv sync --extra test
cd web && npm ci && cd ..
2. 配置模型
cp .env.example .env
至少填下面两组。两组可以先指向同一个模型服务:
# 主 Agent:负责对话和决定下一步做什么
MAIN_AGENT_BASE_URL=https://你的服务/v1
MAIN_AGENT_API_KEY=你的密钥
MAIN_AGENT_MODEL=你的模型名
MAIN_AGENT_TIMEOUT_SECONDS=120
# 专家默认通道:负责简历识别、匹配、邮件处理和面试准备
RESUME_ANALYSIS_AGENT_BASE_URL=https://你的服务/v1
RESUME_ANALYSIS_AGENT_API_KEY=你的密钥
RESUME_ANALYSIS_AGENT_MODEL=你的模型名
BASE_URL 填 /v1 基地址或完整的 /chat/completions 地址都可以。.env 已被 Git 忽略,不要提交真实密钥。
模型选择建议:
-
QwenCloud:优先推荐。
qwen3.7-plus适合主 Agent、岗位分析、简历匹配和简历定制;qwen3.8-flash适合摘要、邮件识别和模拟面试。QwenCloud 提供 OpenAI 兼容的 Chat Completions 接口,并支持函数调用和结构化数据。官方模型选择 · 官方兼容接口 -
DeepSeek:
deepseek-flash适合成本敏感的普通对话、摘要和结构化分析;官方接口支持 JSON Output 和 Tool Calls。官方模型与能力 主 Agent 用 DeepSeek 时必须再加下面两行;不加时请求按 Qwen 的格式发出,模型调用工具后的下一次请求会被 DeepSeek 以 400 拒绝:MAIN_AGENT_VENDOR=deepseek MAIN_AGENT_ENABLE_THINKING=false -
岗位调研:继续使用 GPT 通道,因为当前实现依赖 Responses
web_search;不要把普通 Chat Completions 模型直接填到这个通道。
对话摘要是独立的 Agent worker,但默认复用 MAIN_AGENT_* 的连接;只有需要单独模型时才填写完整配置:
CONVERSATION_SUMMARY_AGENT_BASE_URL=${MAIN_AGENT_BASE_URL}
CONVERSATION_SUMMARY_AGENT_API_KEY=${MAIN_AGENT_API_KEY}
CONVERSATION_SUMMARY_AGENT_MODEL=更快的摘要模型
CONVERSATION_SUMMARY_AGENT_TIMEOUT_SECONDS=30
CONVERSATION_SUMMARY_AGENT_MAX_INPUT_TOKENS=32000
CONVERSATION_SUMMARY_AGENT_MAX_OUTPUT_TOKENS=1200
CONVERSATION_SUMMARY_AGENT_CONTEXT_WINDOW_TOKENS=65536
摘要 worker 在启动时校验输入预算能否容纳一条最大长度消息、活动事实及请求框架;明显偏小的配置直接报配置错误,不等后台进入重试。默认 32000 保持有效。
当前保守下限约 27.7k token(随模型名、请求框架及输出配置略有变化)。未单独配置摘要 worker 时,它沿用 MAIN_AGENT_MAX_INPUT_TOKENS;主模型输入上限降到 24k、16k 或其他低于该下限的值,会导致启动报配置错误,即使普通中文样本实际放得下。此时需完整设置独立的 CONVERSATION_SUMMARY_AGENT_BASE_URL、_API_KEY、_MODEL,并将 _MAX_INPUT_TOKENS 设为 32000 或更高,同时满足该 worker 的输入加输出不超过 _CONTEXT_WINDOW_TOKENS。
这是最低容量校验:已有摘要的其他字段仍占预算,不能保证任意最大长度内容组合都能放下。
岗位调研单独配置为支持 Responses web_search 的 GPT 通道;这里必须填写完整的 URL、密钥和模型:
JOB_RESEARCH_AGENT_BASE_URL=${RESUME_ANALYSIS_AGENT_BASE_URL}
JOB_RESEARCH_AGENT_API_KEY=${RESUME_ANALYSIS_AGENT_API_KEY}
JOB_RESEARCH_AGENT_MODEL=支持web_search的GPT模型
JOB_RESEARCH_AGENT_TIMEOUT_SECONDS=120
模拟面试和简历定制可以单独换模型。简单配置只覆盖模型名和超时,复用 RESUME_ANALYSIS_AGENT_* 的连接:
MOCK_INTERVIEW_AGENT_MODEL=更快的模型名
MOCK_INTERVIEW_AGENT_TIMEOUT_SECONDS=60
MOCK_INTERVIEW_AGENT_PROTOCOL=chat_completions
RESUME_TAILORING_AGENT_MODEL=更强的模型名
RESUME_TAILORING_AGENT_TIMEOUT_SECONDS=120
*_PROTOCOL 可选 responses 或 chat_completions。使用只提供 OpenAI 兼容 Chat Completions 的模型服务时要填 chat_completions,否则请求可能一直等到超时。若某个能力要改用其他模型服务,则为该通道一起填写 BASE_URL、API_KEY、MODEL;不要只填半组:
RESUME_TAILORING_AGENT_BASE_URL=https://官方模型服务/v1
RESUME_TAILORING_AGENT_API_KEY=另一把密钥
RESUME_TAILORING_AGENT_MODEL=官方快速模型
不填写时仍使用 RESUME_ANALYSIS_AGENT_*。当前尚未单独拆出的能力是简历解析、岗位分析、简历匹配、邮件识别和面试准备,它们继续共用这个专家默认通道。
配置关系可以概括为:
| 配置前缀 | 运行时职责 | 是否必须独立填写 |
|---|---|---|
MAIN_AGENT_* | 对话、工具选择和流程编排 | 必须 |
RESUME_ANALYSIS_AGENT_* | 专家默认通道 | 必须 |
CONVERSATION_SUMMARY_AGENT_* | 对话摘要 worker | 可选,默认复用主 Agent |
JOB_RESEARCH_AGENT_* | 岗位/公司调研和网页搜索 | 可选;启用时需完整填写 |
MOCK_INTERVIEW_AGENT_* | 模拟面试 | 可选,默认复用专家通道 |
RESUME_TAILORING_AGENT_* | 简历定制、定稿和评审 | 可选,默认复用专家通道 |
CAREER_EMBEDDING_* 是可选的向量检索服务,不是对话 Agent;CONTEXT_* 只是上下文压缩参数。
3. 创建本地访问密钥
两条命令要用同一个 --user-id,这样网页和浏览器扩展保存的数据才属于同一个用户:
uv run career-agent api-keys issue --user-id local-user --name local-web \
--scope workspace:read --scope workspace:write \
--scope chat:write --scope settings:write --no-expiry
uv run career-agent api-keys issue --user-id local-user --name boss-extension \
--scope capture:write --no-expiry
每条命令只显示一次 secret。把它们分别填进 web/.env:
CAREER_AGENT_WEB_API_KEY=第一条命令返回的_secret
VITE_CAPTURE_API_KEY=第二条命令返回的_secret
4. 启动
uv run career-agent-api # 终端一,后端默认 http://127.0.0.1:8000
cd web && npm run dev # 终端二,前端默认 http://127.0.0.1:5173
浏览器打开 http://127.0.0.1:5173 即可使用。curl http://127.0.0.1:8000/ready 返回 503 时,按响应里提示的配置错误检查 .env。修改 web/.env 后需要重启前端。
5. 安装 BOSS 岗位收藏扩展(可选)
- 在 Chrome
// HOW IT'S BUILT
// FILES
2 files shipped with this skill. Inline text is the catalog copy. The upstream link is the source repository.
skills/job-research/SKILL.md6.1 KBupstream ↗
--- name: job-research description: Research public company, product-line, business, market, and competitor context for one saved job when the user explicitly requests it. Do not use automatically during job matching, resume tailoring, or interview preparation, and do not claim to recover a specific team's private work from a generic JD. --- # Job Research Research public context that helps the user understand a saved job. Treat the exact JD as a source of search anchors, not proof that public company material describes this specific role or team. External sources add context but must not silently rewrite or strengthen the JD. This is an optional, user-requested activity. Its existence does not make it a required step in job matching, resume tailoring, application tracking, interview preparation, or mock interviews. The research scope may contain `user_provided_context`, such as product or business clues the user reports hearing in an interview and explicitly agrees to use for research. Treat it only as an unverified search lead. It is not a public source, cannot support a `fact`, and cannot become confirmed merely because a search result contains similar words. Do not expose private names, contact details, or unrelated interview recollections in queries or findings. ## Research priorities Spend limited source capacity in this order: 1. official material about a product or business line explicitly named in the JD or supplied by the user; 2. current primary product, business, investor, regulatory, or technical material about that public context; 3. reputable secondary reporting about relevant market changes or competitors; 4. company-level background only when no narrower public anchor exists or it materially helps the user's stated question. Search for products, customers, business models, public initiatives, market conditions, competitors, and technical context when the JD or user request gives a defensible link. Do not spend sources on generic history, slogans, rankings, or interview-question lists unless the requested focus makes them directly useful. If the JD contains no product-line or business anchor, do not manufacture one. Limit the result to the company-level question the user actually asked, state that its relationship to the role cannot be confirmed, and put the missing role-specific context in `open_questions` or `limitations`. ## Chinese-market sources Generic web search underserves Chinese employers: the useful primary material is concentrated in places a broad query rarely surfaces. When the company is Chinese, spend source capacity on these before falling back to general results. | Question | Where the primary material is | |---|---| | Products, architecture, technical direction | the company's own engineering blog, its open-source repositories, conference talks | | Business lines, funding, ownership, scale | 招股书, 年报, 投资方公告, 工商信息 (天眼查, 企查查) | | Market position, competitors, recent moves | 36氪, 晚点LatePost, 界面新闻, 第一财经, and the company's own announcements | | Regulatory or licensing constraints | the issuing body's own publication | Prefer the company's own material and the issuing body's own publication over reporting about them. Chinese tech media frequently republishes the same wire copy; three outlets carrying one story are one source, not three, and citing them as three overstates the support behind a finding. ### Self-reported content is not a primary source Anonymous or pseudonymous posts on career and social platforms — 脉脉, 牛客网, 知乎, 小红书, CSDN, and similar — are individual recollections. They may not support a `fact` under any circumstances. At most they support an `inference` with `low` confidence, phrased as what some people report rather than as what is the case, and only when several independent posts agree and nothing primary contradicts them. ### What belongs to interview preparation instead Compensation bands, interview loops, question banks, and hiring-process timelines are out of scope here even when such posts are easy to find. Job research answers what the company does and where it stands; the material a candidate needs to walk into a specific interview is a different request with a different consumer. Do not widen a company-research run into it, and do not report those numbers as company facts. ## Evidence boundaries - A `fact` states only what cited sources support. - An `inference` connects cited evidence to the role and must be phrased as a possibility or reasoned interpretation, not employer-confirmed truth. - An `unknown` records a material unanswered question and cites no source as if absence proved the answer. - Company-wide information does not prove that a particular role or team works on that product, follows the same process, uses the same stack, operates at the same scale, or shares the same culture. - Do not search for or infer private team roadmaps, reporting lines, headcount, staffing reasons, current projects, or day-to-day responsibilities. - Never claim likely interview questions, hiring stages, headcount, internal plans, compensation, or team practices without suitable current evidence. - Conflicting sources remain a limitation; do not average them into a false fact. ## Sources Use only sources actually returned by web research in this run. Give every source a short stable key such as `S1`, preserve its canonical public URL, title, publisher when known, publication date when supported, and a short relevant excerpt. The excerpt is evidence, not a place to copy an article. Every `fact` and `inference` must cite at least one returned source key. Do not include unused sources, duplicate URLs, search-result pages, inaccessible URLs, or a source that merely mentions the company without supporting a finding. ## Output quality Focus the summary on the public company, product, and business context the user asked for. Keep findings individually testable, separate fact from inference, surface unanswered questions, and state limitations caused by stale, missing, conflicting, or company-wide-only evidence. Return only the configured structured response.README.mdupstream ↗
Contents are not inlined in the catalog. Open the upstream file.
Plain-text copy, no login: /skill-bundles/low-hands-mycareer-skills-job-research.md