Career Job Hunting

Career Job Hunting with Hermes Agent

Overview

This skill covers the end-to-end workflow for using Hermes Agent to find, track, and apply for jobs/internships — optimized for the Chinese job market (AI/tech roles). The user (冯周杰 / Publieople) is a 大二 AI 专业学生 at 上海中侨职业技术大学, targeting AI toolchain development / Agent application / MLOps roles.

Components

1. Resume Setup (visiky-resume-setup)

The user's resume data lives in publieople/publieople/resume.json on GitHub. The online editor preview is at https://visiky.github.io/resume?user=publieople&branch=master.

See the visiky-resume-setup skill for full schema, deployment, and debugging.

Key fields to update (last updated 2026-06-04):

• profile.positionTitle — "AI 工具链开发与智能体应用 · 自动化效率工程师"
• educationList — 上海中侨职业技术大学, 人工智能（主修：机器学习、深度学习、计算机视觉、自然语言处理、数据结构与算法）, 2024-至今
• workExpList — 4 bullet points with quantification (GPU集群管理 + AI服务栈 + 培训 + 知识传承)
• projectList — 4 projects (AI Agent生态/产研社/技术博客/AI学习小组)，均补充量化数据
• awardList — 4 awards, ordered by actual prestige: 数维杯>中青杯>五一建模>工行杯
• aboutme — 精简为3行，无MBTI，突出AI Agent闭环经验

2. Job Search (mcp-jobs MCP Integration)

Tool: mergedao/mcp-jobs v1.4.0 — aggregates jobs from 猎聘, BOSS直聘, 智联, 51job via Playwright crawling.

Registration in ~/.hermes/config.yaml:

mcp_servers:
  jobs:
    command: "node"
    args: ["/home/po/.npm-global/lib/node_modules/mcp-jobs/dist/mcp.js"]

Available MCP tools (after Hermes restart):

• mcp_jobs_mcp_search_job — keyword-based job search across platforms
• mcp_jobs_mcp_job_detail — fetch details from a specific job URL

Usage patterns (directly in conversation):

• "搜上海AI实习岗" → search_job(keyword="AI", city="上海")
• "搜北京Python后端薪资15k以上" → search_job with filters
• "看看这个职位的详情" → job_detail with URL from search results

See references/mcp-jobs-deployment.md for install and troubleshooting details.

3. Tool Deployment — Detailed Installation & Integration

For each tool's detailed reference see references/tool-details.md.

3a. mcp-jobs — Zero-Config Job Aggregator

Type: Aggregator (search only) — crawl BOSS, 猎聘, 智联, 51job Tech: Node.js + Playwright + MCP | Risk: Low

Prerequisites: Node.js v18+, npm, Playwright Chromium (npx playwright install chromium) Standalone: npx -y mcp-jobs Hermes Integration (see references/mcp-jobs-deployment.md):

mcp_servers:
  jobs:
    command: "npx"
    args: ["-y", "mcp-jobs"]
    timeout: 180
    connect_timeout: 60

Available tools (after Hermes restart): mcp_jobs_mcp_search_job, mcp_jobs_mcp_job_detail

Known pitfalls:

• npx -y mcp-jobs has no bin field — must use the working config above; the node/direct dist/mcp.js path is a fallback
• v1.4.0 is ~10 months old — platform DOM may have changed
• Proxy: MCP subprocess inherits filtered env from Hermes config, not shell env. If behind http://127.0.0.1:7890, add env: block to the server config
• First run downloads Playwright browser — may take 30-60s

3b. JobClaw — LLM-Powered Smart Apply

Type: Auto-applier (precision, LLM-matched) — BOSS + LinkedIn Tech: Python + Playwright + LLM scoring | Risk: Medium

Prerequisites: Python 3.11+, Playwright (playwright install)

git clone https://github.com/slothsheepking/jobclaw.git
cd jobclaw
pip install -e .
playwright install
jobclaw login
# Edit profiles/me.yaml with skills, salary, city
jobclaw run

LLM auth (priority order): Claude OAuth (free) > Anthropic API key > OpenAI API key Notification: Telegram bot / Discord webhook Anti-detection: Randomized apply delays, HR inactivity filter (7d default) Proxy: HTTP_PROXY / HTTPS_PROXY env vars

Known pitfalls:

• Cookie management: cookies in ~/.jobclaw/cookies/ expire; run jobclaw login --check to verify
• LinkedIn requires separate cookie login
• LLM auth priority means no API key needed if Claude Code is set up — uses subscription token

3c. get_jobs — Mass Batch Apply

Type: Auto-applier (volume) — BOSS (150/day), 猎聘, 前程无忧, 智联 Tech: Java + Gradle + ChromeDriver | Risk: Medium-High (account restrictions)

Prerequisites: JDK 21, Maven/Gradle, ChromeDriver matching Chrome version

git clone https://github.com/loks666/get_jobs.git
cd get_jobs
./gradlew build

Platform limits:

• BOSS: 150 hellos/day hard cap; auto-blacklists "not suitable" companies
• 猎聘: unlimited hellos, must bind WeChat
• 前程无忧: degraded, limited applies
• 智联招聘: ~100 applies, broken, not recommended

Known pitfalls:

• Windows-oriented: hardcoded chromedriver.exe paths; ChromeDriver must match Chrome version exactly
• WSL complication: Chrome must run headless; ChromeDriver path must be Linux-adjusted
• Account risk: BOSS auto-apply patterns can trigger account restrictions — use a secondary account
• .env config: Requires API key for LLM greeting generation (OPENAI_API_KEY or proxy)
• Cost: ~$0.06/day with gpt-5-nano

3d. GeekGeekRun (牛人快跑) — Desktop GUI

Type: Auto-applier (desktop UI) — BOSS only Tech: Puppeteer + Electron | Risk: Low

Packages: Windows (.exe), Linux (.deb), macOS (.dmg) Features: Auto-chat, read-without-reply retry, zombie job cleanup, LLM template editor, config templates Repo: github.com/geekgeekrun/geekgeekrun (857 commits, 27 releases)

4. Deployment Decision Tree

Want to just browse the market?
  └─ mcp-jobs (integrate into Hermes, search by conversation)

Want precision-targeted apply (high match rate)?
  └─ JobClaw (Python, LLM scoring, anti-detection)

Want mass batch apply (volume over precision)?
  └─ get_jobs (Java, multi-platform, cap-aware)

Want a simple GUI without command line?
  └─ GeekGeekRun (Desktop app, BOSS only)

5. Environment Notes (This User — Verified May 2026)

• Shell: Arch Linux via WSL2
• Node.js: v26.1.0 ✅ | npm: 11.14.1 ✅
• Playwright: 1.59.0 (chromium-1217, chromium_headless_shell-1217) ✅
• Python: 3.11.14 ✅
• Java: NOT installed ❌ (needed for get_jobs)
• Proxy: http://127.0.0.1:7890 (for HTTP/HTTPS traffic)
• HF Mirror: HF_ENDPOINT=https://hf-mirror.com
• Resume repo: publieople/publieople with resume.json (visiky/resume format)

6. Application Tracking (Notion)

Set up a Notion database with columns:

• Company, Position, Platform, URL
• Status (已投递 / 已读 / 约面 / 一面 / 二面 / Offer / 感谢信)
• Date applied, Last follow-up, Notes
• Resume version used

Use Hermes' notion skill to interact with this database.

Quick Start for New Sessions

1. Resume: Ensure publieople/publieople/resume.json is up to date → push to GitHub
2. Search: Use mcp-jobs tools (requires Hermes restart after config change) — "帮我搜上海AI实习岗"
3. Track: Log interesting matches in Notion
4. Apply: For high-match roles, draft custom greeting and submit manually or via JobClaw

Job Hunting Timeline (2026)

Key strategy: 福州暑假无大厂AI机会 → 7-8月打磨简历 + 线上调研 → 9月上海日常实习入职。 日常实习全年可投、竞争小于暑期实习、可拉长3-6个月含金量更高。

Pitfalls

• mcp-jobs may return empty: v1.4.0 is ~10 months old and platform DOM changes can break it. Fallback: use web_search with site-specific queries (site:shixiseng.com, site:zhipin.com, 搜索 "上海 AI Agent 实习 2026"). web_search returns richer, more current results than mcp-jobs in many cases.
• mcp-jobs has no bin field: npx -y mcp-jobs fails. Must install globally and use node with direct dist/mcp.js path. See references/mcp-jobs-deployment.md.
• Proxy: Behind http://127.0.0.1:7890, Node.js MCP subprocess inherits filtered env from Hermes. If mcp-jobs can't reach target sites, try passing proxy env explicitly:

  env:
    http_proxy: "http://127.0.0.1:7890"
    https_proxy: "http://127.0.0.1:7890"
  

• Playwright browsers: mcp-jobs requires Chromium. Install with npx playwright install chromium if not present.
• get_jobs is Windows-oriented: uses chromedriver.exe and IntelliJ IDEA. Hard to run in WSL without modifications. Requires JDK 21 which is NOT installed.
• BOSS account risk: Auto-apply patterns (get_jobs, JobClaw) can trigger account restrictions. Use a secondary account for volume apply.
• JobClaw cookie management: Cookies in ~/.jobclaw/cookies/ expire; run jobclaw login --check to verify before batch runs.
• Session recall tool is unreliable: do not use session_search to find past internship discussions. Memory and skills carry the durable state.

Phase 0: Reconcile User Intent (Before Any Planning)

Always do this before writing a plan, picking companies, or editing the resume. A user saying "let's discuss my internship plan" is NOT permission to start drafting — it's permission to discover what the plan should be.

Pull signals from every available source and check them against each other:

Notion API 2025-09 pitfalls when doing this:

• POST /v1/search filter.value MUST be "data_source", not "database" (returns 400 otherwise — older tutorials are wrong)
• Sort + no query = 400. Omit sort if you have no query.
• 404 on /data_sources/{id}/query does NOT mean "doesn't exist" — it means "integration not shared with this database". Try a different data_source.

Then surface the contradictions directly. In this user's case, the Notion 求职意向 table contained 28 WLB-first foreign-tech companies (VMware/AWS/NVIDIA/Microsoft/Apple/...) while resume.json said "AI 工具链开发与智能体应用 · 自动化效率工程师". Both are real, but they don't match — and HR will read whichever is on the resume. Don't paper over this with a "best of both worlds" recommendation; ask the user to pick.

Use the Reconnaissance Report format when reporting back (see references/reconnaissance-report-template.md):

1. 已确认画像 (identity, skills, projects, honors) — concrete facts only
2. 求职意向表的真正信号 — what the company list + reasons reveal
3. 4 个尖锐问题 — explicit mismatches between stated position and inferred direction
4. 5 个事实问题 (A-E) — what the user must answer before planning resumes

Only after the user answers the A-E questions does planning begin. Never write a resume, company list, or timeline until intent is aligned.

Resume Optimization Rules (from 2026-06-04 session)

When helping the user optimize their resume:

1. Never guess competition prestige — verify via web_search before ranking. 校级 awards go last.
2. Don't add projects the user hasn't vouched for — ask first. The user knows which projects are solid vs. "不能细看".
3. University courses may not matter — the user considers self-taught skills more important than formal coursework. Ask before adding.
4. Quantify everything possible — 子弹点 = 动词 + 事项 + 数据. "100+ Skills模块" beats "沉淀百余个".
5. Personal evaluation stays short — 3 lines max, no MBTI, no filler.
6. Section title "更多信息" should be renamed to something concrete like "竞赛获奖" if that's what it contains.

Related Skills

• visiky-resume-setup — online resume JSON schema, deployment, and debugging
• native-mcp — general MCP server configuration for Hermes (includes the no-bin-field fix for mcp-jobs)
• notion — application tracking database management

References

• references/mcp-jobs-deployment.md — mcp-jobs install, troubleshooting, and MCP integration details
• references/tool-details.md — detailed tool comparisons: mcp-jobs v1.4.0 deps, get_jobs platform limits, JobClaw LLM auth, GeekGeekRun features, environment verification
• references/job-tools-landscape.md — broader auto-apply tool landscape (NEAR agent market, OpenOutreach, MoneyPrinterTurbo, etc.)
• references/reconnaissance-report-template.md — Phase 0 user-intent alignment: 4-source cross-check, 4-section report format, A-E gate questions