news
Reviews job application updates from Gmail, LinkedIn and platforms. Prepares drafts, presents executive summary by priority, validates with the user and sends replies.
galiprandi · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add galiprandi/job-seeker --skill news技能说明
name: news description: Reviews job application updates from Gmail, LinkedIn and platforms. Prepares drafts, presents executive summary by priority, validates with the user and sends replies. trigger: news
News
Trigger
Keyword: news
The user says news (or variants: "updates", "check", "any updates") and the full review routine is automatically triggered. No further instructions needed — the agent executes the entire flow from start to finish.
Also runs in parallel when the user launches an application.
Flow
0. Pre-flight
- Verify active LinkedIn and Gmail sessions. If session closed → open browser with wrapper (see AGENTS.md "Browser session"):
node scripts/browser.js open <url> --headed(Gold Rule 5) → notify user → wait for confirmation - Browser: always use
node scripts/browser.jsfor open/close/goto. See AGENTS.md "Browser session" for details. Never callplaywright-cli opendirectly, never open Chrome directly - Load active preferences (see
memoryskill):node scripts/db.js "SELECT category, key, value, confidence, source FROM preferences WHERE user_id = <user_id> AND status = 'active' ORDER BY category, key" - Load strategy (see AGENTS.md "Strategy levels"):
Respect:node scripts/db.js "SELECT data->'strategy' AS strategy FROM users WHERE id = <user_id>"follow_up_days(days before sending follow-up),cold_outreach(whether to send cold messages to recruiters). Ifnewsnot insources_active, warn the user
1. Collect updates (in parallel)
Parallelization strategy: when subagents are available, dispatch background subagents (subagent_general) per source to collect updates simultaneously. Each subagent returns a structured list of items (sender, subject, snippet, category guess, action items, scheduling links if any). The main agent then merges and classifies. If subagents are not available (e.g: single-session constraint), fall back to sequential collection.
Subagent dispatch pattern:
┌─────────────────────────────────────────────────────┐
│ Main agent (orchestrator) │
│ - Loads preferences, strategy, availability │
│ - Dispatches subagents in parallel │
│ - Merges results, classifies, presents summary │
├─────────────────────────────────────────────────────┤
│ Subagent A (Gmail) Subagent B (LinkedIn) │
│ - Inbox unread - Messages unread │
│ - Job Alerts folder - Notifications │
│ - Extract sched links - Saved Jobs │
│ - Returns JSON list - Returns JSON list │
├─────────────────────────────────────────────────────┤
│ Subagent C (DB) Subagent D (Sched links) │
│ - Pending follow-ups - Opens each Calendly/SR │
│ - Pipeline stages - Filters by availability │
│ - Returns JSON list - Returns slot table │
└─────────────────────────────────────────────────────┘
Important: subagents share the same browser. To avoid conflicts, use attached sessions with --session (see AGENTS.md "Parallel execution"):
- Gmail subagent:
node scripts/browser.js attach --session news-gmailthennode scripts/browser.js goto <url> --session news-gmail - LinkedIn subagent:
node scripts/browser.js attach --session news-linkedinthennode scripts/browser.js goto <url> --session news-linkedin - Alternatively, use separate tabs within the same session (
tab-new) if subagents can't use separate sessions - DB subagent doesn't need browser, only
scripts/db.js - Scheduling link subagent:
node scripts/browser.js attach --session news-schedthen opens each link with--session news-sched - When done:
node scripts/browser.js detach --session news-gmail(neverclose— it's ref-counted and would refuse or kill the browser for other agents)
If subagents are NOT available (e.g: tool not supported, single foreground agent constraint), fall back to sequential collection as before. The flow must work in both modes.
Subagent prompt template (adapt per source):
You are a job search assistant. Collect updates from <source> and return a structured list.
Context:
- Last review: <last_review_at>
- User profile: <profile summary from DB>
- Strategy: <strategy level and params>
Instructions:
1. Open <url> using: node scripts/browser.js open <url> (from the repo root directory)
2. <source-specific steps: read unread messages, extract sender/subject/snippet/date>
3. For each item, identify: sender, subject, date, snippet (first 200 chars), category guess (interview/offer/recruiter_new/recruiter_reply/rejected/newsletter/new_job), action items (calendar link? CV requested? form to fill?), and any scheduling URLs
4. Return a markdown table with all items found. Do NOT reply to anything, do NOT archive, do NOT click scheduling links (just extract the URL)
5. Close the tab when done
Return format:
| # | Sender | Subject | Date | Category | Action items | Scheduling URL |
Sources to collect (dispatch as parallel subagents when possible, sequential otherwise):
- Gmail inbox: search for unread emails since last review. Filter: everything related to job search and job sites (recruiters, HR, platforms, newsletters with jobs, application responses). Ignore obvious spam. Save
last_review_atto DB to know since when to search - Gmail
Job Alertsfolder: checkJob Alertslabel (alerts from platforms configured viaradarskill). Classify each alert by fit: Must/Strong/Nice per PROFILE.md. Only present Must and Strong in the summary. Ignore Nice unless user asks to see all - LinkedIn messages: unread messages in inbox. Filter recruiters, HR, application responses
- LinkedIn notifications: application notifications (status changes, recruiter messages)
- LinkedIn Saved Jobs: navigate to
https://www.linkedin.com/my-items/saved-jobs/. For each saved job: check if