Linkedin jobs fetch

LinkedIn Jobs Fetch Skill

Automate fetching saved jobs from https://www.linkedin.com/jobs-tracker/?stage=saved using Playwright with a persisted session stored in ~/.claude/linkedin-session.json.

The working script is bundled with this skill as linkedin_fetch.mjs — don't rewrite it from scratch, it already handles LinkedIn's quirks (see "How the script handles LinkedIn's DOM" below).

Two modes

1. Scrape (default) — fetch all saved jobs as clean JSON, then pass to job-analyzer
2. Add/edit notes (--add-notes) — write or update notes on specific jobs

---

Scrape mode

Step 1 — Ensure playwright + the bundled script are in place

mkdir -p /tmp/pw-runner
ls /tmp/pw-runner/node_modules/playwright 2>/dev/null || (cd /tmp/pw-runner && echo '{"type":"module"}' > package.json && npm install playwright --save && npx playwright install chromium)
cp ~/.claude/skills/linkedin-jobs-fetch/linkedin_fetch.mjs /tmp/pw-runner/linkedin_fetch.mjs

Step 2 — Run the script

cd /tmp/pw-runner && node linkedin_fetch.mjs 2>&1

First run (no session yet): A browser window opens to the LinkedIn login page. Tell the user:

"A browser window has opened. Please log in to LinkedIn. Once you're on the home page (linkedin.com/feed), run this in a new terminal: touch /tmp/linkedin-ready"

Wait for the user to confirm, then the script saves the session automatically. Future runs skip login.

Subsequent runs: Uses ~/.claude/linkedin-session.json — no interaction needed. If the session has expired, the script detects the login/authwall redirect and prints a clear message to delete the session file and re-run.

Step 3 — Parse output and analyze

The script prints a JSON block between === ALL JOBS BY TAB === and === END ===, with each job parsed into title, company, posted, and url:

{ "Saved": [ { "stage": "saved", "id": "123", "title": "...", "company": "... · Oslo", "posted": "Posted 2d ago", "url": "https://www.linkedin.com/jobs/view/123/" } ] }

Extract it and pass to the job-analyzer skill for full analysis.

---

Add/edit notes mode

Two sub-workflows:

A) AI analysis notes — fetch all saved jobs, run job-analyzer on every JD, then write a one-line verdict (rating + reason) as a note on each job card. The user sees the assessment directly in their LinkedIn tracker.

B) Manual notes — user specifies specific jobs and custom text.

Step 1 — Build a notes JSON file

Create /tmp/job-notes.json mapping job URLs to note text (max ~250 chars per note):

{
  "https://www.linkedin.com/jobs/view/1234567890/": "⭐⭐⭐⭐⭐ Strong fit. Domain and seniority align well. Apply.",
  "https://www.linkedin.com/jobs/view/9876543210/": "❌ Target language required. Hard blocker."
}

For the AI-analysis workflow, fetch all jobs first (scrape mode), run job-analyzer on all JDs, then build this JSON with a short verdict per job.

Step 2 — Run with --add-notes flag

cd /tmp/pw-runner && node linkedin_fetch.mjs --add-notes /tmp/job-notes.json 2>&1

For each job ID the script auto-detects which path to use:

• No note yet → clicks the inline "Add note" link in the card's Notes cell.
• Note already exists → there is NO inline link. The script clicks the ⋯ overflow menu button (next to "Apply", aria-label="Overflow menu"), then clicks "Edit note" in the popover that appears.

Both open the same modal; the script fills the textarea and clicks Save. Notes are idempotent — adding to a job that already has a note overwrites it via the edit path.

A verification screenshot of the final tracker state is written to /tmp/linkedin-notes-result.png — read it to confirm the notes landed.

If it crashes mid-run: create a partial JSON with only the remaining jobs and re-run. It's safe to re-run — re-writing/editing an existing note just overwrites it.

---

How the script handles LinkedIn's DOM

• Each card is rendered twice (a tile layout and a table-row layout); only one copy is visible at a given width. The script always targets the visible copy (checks offsetParent + client rects) — never a blind .first(), which would hang ~30s on a hidden element and can close the browser ("stuck"). Targeting is done by tagging the right element with a data-pw-* attribute inside page.evaluate, then acting on it with a Playwright locator (auto-scroll + actionability waits).
• The list virtualizes rows, so the scraper accumulates cards while scrolling incrementally rather than taking a single snapshot.
• Scrape output is parsed, not a raw blob: title / company / posted come from the card's <p> tags, preferring the visible (non-empty innerText) copy.

---

Troubleshooting

• Session expired — the script auto-detects the login/authwall redirect and tells you. Delete ~/.claude/linkedin-session.json and re-run; the login flow will trigger again.
• "Add note" link / overflow menu not found — the job may not be on the saved tab, or LinkedIn changed the DOM. The script walks up from the visible job-link anchor to its card, then looks for either an inline <a>/<button> whose text is exactly "Add note"/"Edit note", or a button[aria-label="Overflow menu"] (for editing existing notes). If LinkedIn renames these, update tagVisibleNoteLink / tagVisibleOverflow in linkedin_fetch.mjs.
• Editing an existing note does nothing / navigates away — the "Edit note" item lives in a body-level popover ([role="menu"]) opened by the ⋯ button; the script clicks it there. If the overflow button moved, re-check openNoteEditor.
• Script hangs waiting for signal — run touch /tmp/linkedin-ready in a terminal once logged in.
• Fewer jobs than expected — the count reflects what's currently on the Saved tab; unsaving a job removes it from the results.