job-search-scheduler

name	job-search-scheduler
description	Two modes: (1) full job scan with deep warm-path detection, and (2) warm-path enrichment on existing scan results. Reads targeting from config, scans LinkedIn + ATS, fetches JDs, detects 1st/2nd-degree warm paths, and writes scan reports + pipeline updates.
triggers	["scan for jobs","js","job search","find matching roles","what jobs match me","scan for jobs today","daily job scan","any new roles","find new roles","run job scan","what's new on the market","enrich warm paths","warm paths","find connections","who knows people at these companies","check my network","find referrals","warm intros for scan","network scan","enrich contacts","find warm leads","run warm path detection","check connections","linkedin people check"]

Job Search Scheduler — Career OS Skill

Task Substrate (v0.25.0+)

$CAREER_GITHUB_REPO is derived from: git -C $CAREER_HOME remote get-url origin | sed 's/.*github.com[:/]//;s/.git$//'

Tasks live in $CAREER_GITHUB_REPO GitHub Issues (canonical source of truth — single inbox for all Cyborg work). Repo of work indicated by repo:* label, NOT by issue location. Cadence indicated by cadence:* label (operational for high-frequency churn; strategic for sprint-scale; meta for trackers). Tier indicated by tier:* label (p1/p2/p3/backlog).

Tasks.md is DEPRECATED as of v0.25.0. See $CAREER_HOME/workspace.manifest.yaml task_routing: section for the full architecture.

This skill reads/writes via:

• gh CLI (universal, all agents)
• github-mcp MCP server (post-restart, when MCP boots — at npx @modelcontextprotocol/server-github)

New role discoveries open as kind:scan-result issues (one per high-priority finding) so cruise-control / mission-control can pick them up:

gh issue create --repo $CAREER_GITHUB_REPO \
  --title "Score+apply: {Company} {Role}" \
  --label "tier:p2,cadence:operational,repo:career-os-data,kind:scan-result" \
  --body "<scan-source URL + JD path + warm-path tag + suggested resume track>"

Scan reports themselves remain markdown files at career-intelligence/projects/job-search/scans/{date}/scan-{HH}-{MM}.md — those are READ artifacts, not tasks.

FIRST-TIME SETUP (new users — run this before your first scan)

Prerequisite: career-intelligence-onboarding must be complete. Say "onboard me to career intelligence" to run it. It creates brain/identity/experience-history.md and career-intelligence/projects/job-search/job-search-config.md from an 8-question interview. Without these files, the scanner has no targeting config and will produce irrelevant results.

Step 1 — Expand job search config After onboarding, say: "set up my job search config" or "expand job search targeting". This reads your preferences from job-search-config.md and creates the full operational config at brain/config/job-search.md — adding ATS direct URLs, LinkedIn search keywords, company tiers, and warm-path settings. This is what the scanner reads on every run.

Step 2 — Add your resume tracks Create a folder Resumes & Cover Letters/ in your workspace. Paste or upload your current resume as one or more files. The file names become your track names (e.g., resume-engineering-leader.md, resume-executive.md). The resume engine auto-detects tracks from filenames. If you have only one resume, name it resume-base.md — the engine will derive tracks from your experience history.

Step 3 — Set up the 6 AM automated scan (optional) Use Claude Code's scheduled task system to run the scanner automatically:

1. Open Claude Code settings → Scheduled Tasks → New Task
2. Set schedule: 0 6 * * 1-5 (6 AM Monday–Friday)
3. Paste the prompt from skills/job-search-scheduler/scan-prompt-v11.md (everything below the ## PROMPT BODY line)
4. Set workspace to your $CAREER_HOME directory The automated scan runs in surface-level mode (no deep People tab navigation). You review the output and run "enrich warm paths" manually for high-priority roles.

What to customize per your situation:

• brain/config/job-search.md → edit Target Roles, Filters, ATS Direct URLs, LinkedIn Search Keywords
• Target level and comp floor come from your onboarding answers — edit them in the config file if your search evolves
• The scanner reads the config on every run; edits take effect immediately

---

Purpose

Two jobs:

1. Scan — discover open roles matching your profile across LinkedIn and ATS career pages
2. Enrich — take an existing scan report and run deep warm-path detection (1st-degree, 2nd-degree, alumni networks) on the roles found

The scan produces a role list. The enrichment produces actionable referral paths. Both write to the same scan directory. Together they form the highest-ROI step in the job search: finding roles AND finding the people who can get you in.

Output Format

Always start your response with:

━━━ Career OS: Job Search ━━━

For enrichment mode:

━━━ Career OS: Warm Path Enrichment ━━━

How to Invoke

Mode 1: Full Scan

Say: "scan for jobs", "any new roles?", "daily job scan", "find matching roles", "what's new on the market", "run job scan"

Runs the complete pipeline: discover → JD fetch → quick filter → warm-path detection → write scan report → pipeline write-back.

Mode 2: Warm-Path Enrichment (on existing scan)

Say: "enrich warm paths", "warm paths", "find connections", "check my network", "find referrals", "warm intros for scan", "network scan", "enrich contacts", "find warm leads", "linkedin people check", "check connections"

Reads the most recent scan report, identifies roles with Cold or surface-level warm tags, and runs deep LinkedIn People tab navigation for 1st-degree, 2nd-degree, and alumni network detection. Writes an enrichment addendum to the same scan date folder.

Mode 3: Automated (Scheduled Task — 6 AM safety net)

Runs via daily-linkedin-job-scan scheduled task. Surface-level warm tagging only (no People tab navigation — too slow for headless runs). Writes scan report for user to review or enrich manually.

---

CROSS-REFERENCES (P9 Coherence)

Artifact	Location	Relationship
Config (targeting)	brain/config/job-search.md	Source of truth for all targeting parameters
Scheduled task prompt	skills/job-search-scheduler/scan-prompt-v11.md	Canonical scanner prompt (v11 — plugin path format). Paste into scheduled-task config.
Mission-control routing	skills/mission-control/SKILL.md	Routes scan + enrich triggers to this skill
Job-match-scorer skill	skills/job-match-scorer/SKILL.md	Receives scan output for deep scoring
Network-intelligence skill	skills/network-intelligence/SKILL.md	Deep contact analysis for individual companies

If you change targeting → edit config only (both consumers read it). If you change scan logic → update this SKILL.md AND the scheduled task prompt.

---

DATA ARCHITECTURE

Inputs (what the skill reads)

Source	Path	What It Provides
Config	brain/config/job-search.md	Target roles, company tiers, ATS URLs, LinkedIn keywords, filters, skip rules, warm-path settings
JD samples	brain/reference/jd-samples/*.md	Saved JD snapshots for roles already fetched
Pipeline	career-intelligence/projects/job-search/job-pipeline.json	Already Applied table (skip dupes) + Warm Intros table (existing contacts)
Skills matrix	brain/identity/skills-matrix.md	Technology proficiency for quick-filter gap detection
People	network/people/*.md	Contact network for warm-path cross-referencing
Identity	brain/identity/identity.md	Values, philosophy for culture fit
LinkedIn (via Chrome MCP)	Browser navigation	Job feeds, company pages, People tabs, connection data
Previous scan	career-intelligence/projects/job-search/scans/{latest}/*.md	Delta tracking — what changed since last scan

Outputs (what the skill writes)

Output	Path	What It Contains
Scan report	career-intelligence/projects/job-search/scans/{YYYY-MM-DD}/scan-{HH}-{MM}.md	Roles found, warm tags, JD quality, verify queue
Enrichment report	career-intelligence/projects/job-search/scans/{YYYY-MM-DD}/enrich-{HH}-{MM}.md	Deep warm-path results for roles from latest scan
JD snapshots	brain/reference/jd-samples/{company}-{role-slug}.md	Fetched JD text (permanent — URLs die, snapshots don't)
Task updates	GitHub Issues $CAREER_GITHUB_REPO	New kind:scan-result issues opened per high-priority role
Pipeline updates	career-intelligence/projects/job-search/job-pipeline.json	New "Ready to Apply" entries

---

BEHAVIOR: Mode Selection (Automatic)

When the skill is invoked, determine mode from the trigger:

Trigger matches	Mode	Action
Scan triggers ("scan for jobs", "any new roles", etc.)	Full Scan	Run complete pipeline (Steps 1-5)
Enrich triggers ("enrich warm paths", "warm paths", etc.)	Enrichment Only	Skip discovery, run warm-path enrichment on latest scan (Step 3 only)
"scan and enrich" or "full scan with warm paths"	Full Scan + Enrichment	Run complete pipeline including deep warm paths

---

BEHAVIOR: Full Scan (Mode 1)

Step 1: Load Context

Read from config file (brain/config/job-search.md):

• Target roles, company tiers, ATS URLs, LinkedIn keywords, filters
• Skip rules, structural gap auto-skip rules
• Resume track mapping

Read pipeline + people files:

• Already Applied table → skip exact role+company pairs
• COUNT the rows yourself — do NOT trust header count
• Warm Intros table → for warm-path tagging
• People directory → company names for tagging

Read most recent scan file → for delta tracking.

Step 2: Scan Discovery Sources

Source A: LinkedIn (Chrome MCP required)

If Chrome MCP unavailable: skip Source A entirely. Log it. Run Source B only.

A1. Recommended feed: Navigate linkedin.com/jobs → capture first 20 matching roles from "Recommended for you."

A2. Keyword searches: Run each keyword from config's LinkedIn Search Keywords section at linkedin.com/jobs/search/ with filters from config's LinkedIn Search Filters section. First 15 results per search. Dedup.

A3. Connection noting: While on each LinkedIn job page, if LinkedIn shows "X connections at this company" or names 1st-degree contacts, note them. Do NOT navigate to People tabs in this step (that's Step 3).

Fallback: If Chrome MCP unavailable or LinkedIn logged out, skip and log.

Source B: ATS Career Pages (WebFetch only)

WebFetch each URL from config's ATS Direct URLs table. Look for role titles matching target roles from config. WebFetch is faster and cheaper than Chrome for server-rendered ATS pages.

JS-rendered sites from config's JS-Rendered Sites list are covered by LinkedIn in Source A only.

Step 2b: JD Fetch Gate (REQUIRED before scoring)

For every NEW role found:

1. Extract direct job apply URL (not company page, not LinkedIn search URL)
2. Validate URL pattern per config rules
3. If valid: fetch JD text, save to brain/reference/jd-samples/, tag ✅ JD
4. If LinkedIn search/company URL only: tag 🔄 partial
5. If 404 / "Job not found": mark CLOSED, exclude
6. If cannot fetch (JS, login wall): add to Verify Queue, tag ⚠️ title-only

JD URL validation rules:

URL Pattern	Valid?	Action
https://jobs.ashbyhq.com/...	Yes	Fetch JD
https://job-boards.greenhouse.io/...	Yes	Fetch JD
https://jobs.lever.co/...	Yes	Fetch JD
https://careers.{company}.com/...	Yes	Fetch JD
https://linkedin.com/company/...	No	Company page — click through
https://linkedin.com/jobs/search/...	No	Search results — not a JD URL

Rescore notification (WO-035): After fetching JDs, check if any newly-fetched JD matches a role already in the match tracker with ⚠️ title-only Quality. If so, append a note to the scan report:

⚠️ {N} roles now have JDs that were previously scored title-only — run "rescore queue" to update scores

This is a notification only — do not auto-trigger rescoring (token cost is Opus-level, user decides when to spend it).

Step 2c: Surface-Level Warm Tagging

For every role found, check company name against:

• Warm Intros table in pipeline
• People directory

Tag with strongest match (priority: 🔥 ACTIVE > ✅ REFERRAL > ✅ Resume sent > other). Note LinkedIn connection counts from A3.

This is the same level the automated 6 AM scan does. Deep detection is Step 3.

Step 3: Deep Warm-Path Detection (Manual Scan Only)

This step only runs when the user invokes the skill manually (not the scheduled task).

For every NEW role found with ≥80% quick filter score, extract connection data from the company's LinkedIn People tab.

Browser Extraction Strategy (WO-037: DOM/JS over screenshots)

Primary method: DOM/JS extraction (10-20x cheaper than screenshots)

1. Navigate directly to the People tab URL: https://www.linkedin.com/company/{slug}/people/ This skips the "navigate to company page → click People tab" round-trip.

2. Extract via JavaScript (javascript_tool or equivalent JS execution): Query DOM for people cards — name, title, connection degree, mutual connections. Do NOT take screenshots to visually parse the page.

3. Filter via URL parameters instead of click-filter-screenshot:

   • 1st connections: /company/{slug}/people/?facetNetwork=%5B"F"%5D
   • 2nd connections: /company/{slug}/people/?facetNetwork=%5B"S"%5D
   • Alumni: /company/{slug}/people/?keywords=Berkeley

4. Rate limiting: Wait 2-3 seconds between companies to avoid LinkedIn throttling.

Fallback: If JS extraction fails (empty results, CAPTCHA, DOM change):

• Log the failure with the specific selector that broke
• Fall back to screenshot-based extraction for that company only
• Surface warning: "DOM extraction failed for {Company} — falling back to screenshot"

3a: 1st-Degree Detection

• Navigate to People tab filtered by 1st connections (URL param or JS filter)
• Extract via DOM: name, title, relationship context
• Cross-reference against network/people/*.md
• Cross-reference against pipeline Warm Intros table

3b: 2nd-Degree Detection (HIGH VALUE)

• Navigate to People tab filtered by 2nd connections (URL param or JS filter)
• For each 2nd-degree contact: extract WHO the mutual connection is from DOM
• If mutual connection is in network/people/*.md → actionable intro path
• Record: 2nd-degree name + title, mutual connection name, relationship strength

2nd-degree scoring:

Mutual Connection Type	Strength	Action
In people/ with 🔥 ACTIVE status	Very strong	Ask mutual for direct intro
In people/ with any status	Strong	Reach out to mutual, mention shared connection
Berkeley Haas EMBA cohort	Strong	Alumni warm outreach
LinkedIn 1st but not in people/	Medium	Connect first, then ask
Poker gang / social circle	Very strong	Text/WhatsApp ask

3c: Alumni Network Detection

• Navigate to People tab with ?keywords=Berkeley (or "Haas")
• Extract count and names via DOM
• 3+ alumni = alumni network tag (medium strength)

3d: Score Adjustments

• 1st-degree referral path: +5% to match score
• 2nd-degree via known contact: +3% to match score
• Alumni network (3+): +2% to match score
• Cold: no adjustment

3e: Chain to network-intelligence

For roles scoring ≥85% with warm paths found, suggest: "who do I know at [Company]" → routes to network-intelligence for deep contact analysis and outreach draft.

Step 4: Quick Filter + Job-Match-Scorer Handoff

Apply quick filter from config to disqualify obvious mismatches. Deep 6-category scoring is handled by the job-match-scorer skill.

Factor	Weight
Role level fit	20%
Company tier	10%
Recency	5%
Salary fit	5%

If quick filter < 50% → SKIP.

After scan completes, suggest: "score the latest scan" to invoke job-match-scorer.

Step 5: Write Outputs

Scan report (career-intelligence/projects/job-search/scans/{YYYY-MM-DD}/scan-{HH}-{MM}.md):

# Job Scan — {date} {time}

## Search Mode: {from config}

**Candidate:** {read from brain/identity/experience-history.md → `who:` frontmatter field}
**Scan type:** Manual scan with deep warm-path detection
**Reference:** Previous scan {date} {time}

---

## Pre-Scan Check
Already Applied: {N} rows counted (header said {M})
Last scan: {date} {time} — {path}
People files: {N} contacts loaded

---

## Summary
| Source | Scanned | New | Skipped | Dead |
|--------|---------|-----|---------|------|

## New Roles
| # | Company | Title | Location | Track | Warm Path | JD Quality | Source | URL |

## Warm Paths Found
| Company | Role | Contact | Type | Strength | Action |
|---------|------|---------|------|----------|--------|

## 2nd-Degree Intro Paths (HIGH VALUE)
| Company | Role | 2nd-Degree Contact | Mutual Connection | Strength | Action |
|---------|------|--------------------|-------------------|----------|--------|

## Alumni Networks
| Company | Alumni Count | Notable Names | Action |

## New Companies (not on ATS list)
| Company | Roles | Stage | Add to list? |

## Status Changes (vs last scan)
| Company | Title | Change |

## Verify Queue
| Company | Role | Source URL | JD Quality | Failure Reason | Action |

## No Matching Roles
| Company | Notes |

## Skipped (with reason)
| Company | Role | Reason |

## Next Steps
→ "score the latest scan" — run Opus scoring on new roles
→ "enrich warm paths" — run deep warm-path detection (if not already done)
→ "who do I know at [Company]" — deep network analysis for a specific company
→ "dashboard" — see apply-ready queue

Pipeline write-back:

• New roles → append to "Ready to Apply" section with "Scan {date} — awaiting scoring"
• Do NOT add scores, do NOT modify existing entries, do NOT touch Already Applied

Task updates:

• Open kind:scan-result issues on $CAREER_GITHUB_REPO for high-priority new roles (one issue per role; see Task Substrate section above for command)

---

BEHAVIOR: Warm-Path Enrichment (Mode 2)

When user says "enrich warm paths", "find connections", "check my network", etc.:

Step 1: Find Unenriched Scans

• Find ALL scan files across career-intelligence/projects/job-search/scans/ (any date directory, pattern scan-*.md)
• For each scan, check if a corresponding enrichment file exists:
  • Enrichment files match pattern enrich-*.md in the same date directory
  • A scan is "enriched" if an enrichment file's Scan enriched: field references it
• Collect roles from ALL unenriched scans (not just the most recent)
• Deduplicate by company + role title (same role in multiple scans = enrich once)
• Identify roles tagged Cold or with surface-level warm tags only
• Load people files and pipeline Warm Intros table

If no scans exist: "No scan found. Say 'scan for jobs' to run one first." If all scans already enriched: "All scans have been enriched. Run a new scan first."

Step 2: Deep Warm-Path Detection

For each role from the unenriched scans that needs enrichment (Cold or surface-level only):

• Run Steps 3a-3e from the full scan (1st-degree, 2nd-degree, alumni detection)
• Use DOM/JS extraction strategy (see Step 3 "Browser Extraction Strategy") — not screenshots
• Requires Chrome MCP for LinkedIn People tab navigation

If Chrome MCP unavailable: "Chrome MCP needed for LinkedIn People tab navigation. Run this in a session with Chrome access."

Step 3: Write Enrichment Report

Save to career-intelligence/projects/job-search/scans/{YYYY-MM-DD}/enrich-{HH}-{MM}.md:

# Warm Path Enrichment — {date} {time}

**Scans enriched:** {scan-09-00.md, scan-10-30.md, ...}
**Roles analyzed:** {N total, M after deduplication}
**New warm paths found:** {N}
**2nd-degree intros found:** {N}

## Warm Path Results

### {Company} — {Role}
| Path | Type | Strength | Action |
|------|------|----------|--------|
| {Name} (1st, {Title}) | Direct referral | Very strong | Ask to refer |
| {Name} (2nd via {Mutual}) | 2nd-degree intro | Strong | Ask {Mutual} for intro |
| {N} Haas alumni | Alumni network | Medium | Alumni outreach |

## Priority Actions (sorted by strength)
| # | Company | Contact | Action | Strength |
|---|---------|---------|--------|----------|

## Still Cold (no paths found)
| Company | Role | LinkedIn Connections | Suggestion |
|---------|------|---------------------|------------|

→ "write outreach for [Name]" — draft intro request
→ "who do I know at [Company]" — deeper analysis on specific company

Step 4: Update Scan Report

If the original scan report has warm-path tags, update them with enriched results. Use surgical edits — do not rewrite the entire scan file.

---

BEHAVIOR: First Run (No Config)

When brain/config/job-search.md does NOT exist:

Walk through setup questions (one at a time, P11):

1. Target role titles
2. Job boards to scan
3. Companies to include/exclude
4. Location preferences
5. Salary floor
6. Preferred scan time

Create config file and register scheduled task via schedule skill.

---

BEHAVIOR: Manual Scan with Filters

User can request targeted scans:

• "Scan for Anthropic roles" → filter to one company
• "Find director-level AI roles posted this week" → filter by level + recency
• "Any new roles at my tier 1 companies?" → use company tier from config

---

BEHAVIOR: Expanded Company Discovery (Weekly)

On weekly cadence or when user says "find new companies":

• Group LinkedIn results by company, flag unknowns
• Check qualification criteria from config
• Add qualified companies to pipeline and config's ATS list

---

MISSION CONTROL INTEGRATION

When dashboard loads, check for today's scan:

• Scan exists + enriched: "{N} roles found, {N} warm paths — say 'dashboard' to apply"
• Scan exists, not enriched: "{N} roles found — say 'enrich warm paths' for referral detection"
• Scan stale (>24h): "Job scan overdue — say 'scan for jobs'"
• No scan config: "Set up job scanning → 'scan for jobs'"

---

MULTI-AGENT SAFETY (P15)

• Read pipeline and people files before every scan (another agent may have updated them)
• Surgical edits to pipeline — append rows, never rewrite tables
• Write scan reports as new files (no overwrites)
• Log all changes to handoff doc