job-search

What this skill does

# Multi-Source Job Search

A Claude Code skill for searching jobs across LinkedIn, Hacker News Who's Hiring, and Twitter/X. Scores and ranks results against your saved preferences, highlights network advantages, and saves structured output.

---

## Phase 1: Load Preferences & Parse Overrides

### Step 1: Load Profile

Read `~/.claude-job-profile.json`. Check for the `preferences` key.

**If no preferences found**, say:

> No search preferences found. Run `/job-preferences` first to set your target titles, salary range, and filters.

Then **STOP**. Do not prompt the user to set preferences inline.

**If preferences exist**, load them:

```json
{
  "targetTitles": ["Staff AI Engineer", "Principal ML Engineer"],
  "minBaseSalary": "$250K",
  "remotePreference": "remote only",
  "excludePatterns": ["junior", "associate", "intern", "entry level"],
  "defaultTimeRange": "last week"
}
```

### Step 2: Parse User Overrides

The user may pass overrides when invoking the skill:

- **Time range**: `last week`, `2 weeks`, `month` — overrides `defaultTimeRange`
- **Source filter**: `linkedin`, `hn`, `twitter`, or any combination — limits which sources to query (default: all three)
- **Keywords**: any additional keywords to use alongside `targetTitles`

Display the active search config before proceeding:

> **Search config:**
> - Titles: Staff AI Engineer, Principal ML Engineer
> - Salary floor: $250K
> - Remote: Remote only
> - Exclude: junior, associate, intern, entry level
> - Time range: last week
> - Sources: LinkedIn, HN, Twitter

---

## Phase 2a: LinkedIn Search (Chrome MCP)

### Navigation

1. Get browser context using `tabs_context_mcp`
2. Create a new tab using `tabs_create_mcp`
3. Navigate to `https://www.linkedin.com/jobs/`
4. Verify user is logged in — use `read_page` to check for profile menu. If not logged in, say: "Please log into LinkedIn in this browser tab, then let me know when you're ready." and **wait**.

### Build Search URL

Base URL: `https://www.linkedin.com/jobs/search/`

| Parameter | Purpose | Values |
|-----------|---------|--------|
| `keywords` | Search terms | URL-encoded from targetTitles |
| `location` | Geographic area | From profile location |
| `f_WT` | Work type | `1` (on-site), `2` (hybrid), `3` (remote) |
| `f_TPR` | Time posted | `r604800` (week), `r1209600` (2 weeks), `r2592000` (month) |
| `f_E` | Experience level | `4` (mid-senior), `5` (director), `6` (executive) |
| `f_JIYN` | In your network | `true` |
| `sortBy` | Sort order | `DD` (date) |

Map `remotePreference`:
- "remote only" → `f_WT=3`
- "remote preferred" → `f_WT=2,3`
- "open to hybrid" → `f_WT=2,3`
- "open to all" → omit `f_WT`

Map time range:
- "last week" → `f_TPR=r604800`
- "2 weeks" → `f_TPR=r1209600`
- "month" → `f_TPR=r2592000`

### Extract Job Listings (max 25)

For each job card:

1. Use `read_page` to read job cards from results list
2. Click into job detail page using `computer` with left_click
3. Wait 2-3 seconds for detail page to load
4. Extract: title, company, location, posted date, applicant count, work type, apply method
5. Look for connection indicators — use `find` with query "connections work here" or "connections at"
6. Look for hiring team — use `find` with query "Meet the hiring team" or "hiring manager"
7. If hiring manager found, extract name, title, profile URL
8. Store result, navigate back to results list
9. **2-3 second delay** between each job

### Scrolling for More Results

LinkedIn uses infinite scroll:
1. Use `computer` with action "scroll", scroll_direction "down"
2. Wait 2-3 seconds for new results
3. Use `read_page` to see newly loaded cards
4. Repeat until desired count or no more results

---

## Phase 2b: Hacker News Who's Hiring (Bash + WebSearch)

### Find Current Thread

1. Use `WebSearch` to search for: `"Ask HN: Who is hiring?" site:news.ycombinator.com {current_month} {current_year}`
2. Extract the thread ID from the HN URL in the search results (e.g., `https://news.ycombinator.com/item?id=XXXXXXXX` → `XXXXXXXX`)

If no thread found for the current month, try the previous month. If still nothing, skip HN and report it.

### Fetch Comments via Firebase API

3. Fetch the thread: `curl -s "https://hacker-news.firebaseio.com/v0/item/{THREAD_ID}.json"`
4. Parse the `kids` array — these are top-level comment IDs (job postings)
5. Fetch each comment (up to 50): `curl -s "https://hacker-news.firebaseio.com/v0/item/{COMMENT_ID}.json"`
6. **0.5 second delay** between API calls

### Parse Comments

HN Who's Hiring comments typically follow this format in the `text` field:

```
Company Name | Role Title | Location | Remote | Salary Range
Description text...
Apply: URL
```

For each comment:
- Extract company, title, location, remote status, salary from the first line (pipe-delimited)
- Extract application URL if present (look for "apply" or "http" links)
- Parse salary range if present (look for patterns like `$XXXk-$XXXk`, `$XXX,XXX`)
- Extract the full text as description

### Filter Against Preferences

Skip comments that:
- Don't match any `targetTitles` (fuzzy match — "AI Engineer" matches "Staff AI Engineer")
- Match any `excludePatterns`
- Don't meet `remotePreference` (if "remote only", skip non-remote postings)
- Fall below `minBaseSalary` (if salary is listed)

---

## Phase 2c: Twitter/X Search (Chrome MCP)

### Navigation

1. Use existing browser context (or create new tab)
2. Navigate to `https://x.com/search`
3. Verify logged in — use `read_page` to check for profile avatar or compose button. If not logged in, **skip Twitter entirely** and continue with other sources. Report: "Skipped Twitter — not logged in."

### Build Search Query

Construct a Twitter advanced search query:

```
("hiring" OR "open role" OR "we're hiring" OR "join our team") ("AI engineer" OR "ML engineer" OR "{title1}" OR "{title2}") ("remote") since:YYYY-MM-DD -is:reply
```

Map time range to `since:` date:
- "last week" → 7 days ago
- "2 weeks" → 14 days ago
- "month" → 30 days ago

### Execute Search

4. Navigate to the search URL with the query
5. Wait for results to load (3 seconds)
6. Switch to "Latest" tab if available
7. Extract tweet content, author handle, engagement (likes/retweets), and any URLs (max 20 tweets)
8. **2-3 second delays** between interactions

### Filter Against Preferences

Apply the same filtering as HN — title match, exclude patterns, remote, salary if mentioned.

### Graceful Failure

If Twitter is inaccessible, rate-limited, or not logged in — skip entirely and continue. Report which sources succeeded and failed at the end.

---

## Phase 3: LinkedIn Detail Extraction

For LinkedIn results that passed initial filtering, extract additional network signals:

1. Connection count and names (1st-degree, 2nd-degree, alumni)
2. Hiring manager name, title, profile URL
3. Easy Apply availability
4. Applicant count and posting age

This data feeds into the scoring in Phase 4.

---

## Phase 4: Cross-Source Scoring

Score every result on a 0-100 normalized scale.

### Scoring Rubric

| Category | Points | Applies to |
|----------|--------|-----------|
| Title match (exact vs partial) | 0-20 | All sources |
| Salary meets/exceeds floor | 0-10 | All sources |
| Remote preference match | 0-10 | All sources |
| Recency (newer = higher) | 0-5 | All sources |
| Low competition (<50 applicants) | 0-5 | All sources |
| Hiring manager listed | 20 | LinkedIn only |
| 1st-degree connections | 15 | LinkedIn only |
| 2nd-degree or alumni connections | 10 | LinkedIn only |
| Easy Apply available | 5 | LinkedIn only |
| Salary explicitly listed | 10 | HN only |
| Application URL provided | 10 | HN only |
| High engagement (50+ likes) | 10 | Twitter only |

### Normalization

Max possible per source: LinkedIn 100, HN 80, Twitter 70.

Normalize all scores to 0-100:
- LinkedIn score: `raw_score`
- HN score: `raw_score * (100 / 80)`
- Twitter score: `raw_score * (100 / 70)`

Round to nearest integer. Sort all results by normalized sco