linkedin-profile-audit（linkedin 个人资料审核）

LinkedIn Profile Auto-Audit & Accuracy Correction

Purpose

Catch the lies on your LinkedIn profile—overclaims, fabricated metrics, vague language you missed—before a recruiter or reference check finds them. Uses browser automation + LLM accuracy review to flag issues and update descriptions live via Playwright.

When to Use

• Before a job search — profile needs to be referral-safe and stand up to peer scrutiny
• After AI-assisted rewrites that may have introduced overclaims or fabricated metrics
• After a period of growth when descriptions may no longer reflect the actual role
• Before reference checks or background screening

Prerequisites

• Playwright CDP connection to a live Chrome session logged into LinkedIn
• Position IDs for each experience entry (visible in LinkedIn edit URLs)
• User available to answer targeted clarification questions about each role

Process

Phase 1: Extract All Descriptions

Navigate to each experience edit URL and extract current description text—batch all reads in a single Playwright script to avoid repeated round trips—and write results to a local file for review before making any changes.

https://www.linkedin.com/in/{profile}/edit/forms/position/{position_id}/

Phase 2: Accuracy Audit — Flag by Severity

For each description, flag:

🔴 High — Fix immediately:

• Fabricated or unverifiable metrics (user counts, % improvements, SLA figures)
• Claimed production deployment when work was PoC, design-only, or local
• Sole attribution ("I built / I architected") for work done by a team you advised or directed
• Actions claimed that didn't happen (e.g. "closed a deal" when someone else closed it)

🟡 Medium — Verify with user:

• "Led" language for tandem or shared leadership roles
• Specific feature names or technical details not confirmed by the user
• Partner/vendor relationships described more intimately than reality
• "Founded/launched" language for roles that were inherited or handed over

🟢 Clear — No change needed:

• Self-owned company work (founder has full ownership)
• Confirmed hands-on builds (AI-assisted is still the user's work)
• Verified, measured metrics
• Roles with clear sole ownership

Phase 3: Targeted Clarification

For each flagged item, ask ONE precise question. Don't bundle — get clear answers one role at a time.

Key questions:

• "Did you personally build/deploy this, or did you direct others who did?"
• "Were these metrics measured in production, or estimated/aspirational?"
• "Was there someone above you on the [commercial/technical] side?"
• "What was the actual relationship with [partner] — tool, co-builder, or just conversations?"
• "Did anything go live or to test, or was this architecture/design phase only?"
• "Was this your own idea, or a collaborative extension?"

Phase 4: Batch Corrections

Rewrite flagged descriptions with corrected language. Apply all corrections in a single Playwright batch script.

Key verb substitutions:

Critical Accuracy Principles

• Advisory ≠ Builder — If you guided/directed others, say so. The people who coded it deserve credit, and peers will know the difference.
• Design ≠ Deployment — Never claim production metrics for work that didn't ship. Local PoCs are valuable — describe them honestly.
• Tandem ≠ Led — If someone else owned the non-technical or commercial side, reflect that explicitly.
• AI-assisted is still yours — Using Cursor, Claude, or Copilot to write code you directed and deployed doesn't diminish authorship. No caveat needed.
• PoC/local ≠ Production — "Built a working local proof-of-concept that validated the approach" is strong and honest. "Deployed to 100K users" when nothing went live is a liability.
• Collaboration is a feature — Saying "worked in tandem with the CEO" or "guided the Monash research team" signals leadership *and* honesty. Referees will confirm the accurate version.

Playwright Notes

Work with .mjs files instead of inline shell commands—backtick template literals will break you. Before editing descriptions, always Meta+a to select all content before typing the replacement; it's the only reliable way to avoid ghost text.

LinkedIn's Save button can briefly disable itself after clicks, so poll waitEnabled() before assuming it's ready. Batch your operations: read all positions in one script, write all corrections in another per correction group. CDP connects to the local loopback address only (never a remote host):

ws://127.0.0.1:18800/devtools/browser/{browser_id}

Output

• Corrected LinkedIn descriptions live on profile
• Fact/data log in memory/YYYY-MM-DD.md for future blog post material
• Referral-safe profile that stands up to peer review from former employers and colleagues