Building an AI System That Manages a LinkedIn Profile Without Getting Caught

Giovanni Liguori — Wed, 25 Mar 2026 05:48:21 +0000

For 22 days, I've been running a system that autonomously manages a professional LinkedIn profile. Daily posting, strategic engagement, DM triage, weekly reporting — all orchestrated by Claude Cowork with zero manual intervention.

Today I'm open-sourcing the entire system as a Claude skill. Here's how it works, what went wrong, and what I learned.

The Problem

Building a professional LinkedIn presence takes 3-4 hours daily: content creation, engagement, analytics, DM management. For a solo professional, that's unsustainable.

I needed a system that could do all of this autonomously while maintaining a recognizable, human voice — and without getting flagged as AI.

Architecture: 5-Phase Wizard

The skill isn't a template. It's a guided wizard that builds a personalized automation system:

Phase 1: Identity & Voice
15 questions extract your personal tone of voice: rhetorical patterns, vocabulary preferences, emotional registers, even a profile blacklist. The system doesn't start from "what to post" but from "who you are when you write."

Phase 2: Strategy & Content
A pillar calendar maps 7 days to 7 content types, each with a different emotional register:

Monday: Behind the scenes (curious, vulnerable)
Tuesday: Tool/Workflow (pragmatic, generous)
Wednesday: Hot take (provocative, moral)
Thursday: Case study (proud, specific)
Friday: How-to (didactic, patient)
Saturday: Storytelling (personal, reflective)
Sunday: Soft CTA (direct, confident)

Phase 3: Engagement & Anti-Detection
This is where it gets interesting. See the next section.

Phase 4: Task Plan — Review & Approve
The user sees all 10 scheduled tasks with times, frequencies, and dependencies. Nothing runs without explicit approval.

Phase 5: Create Tasks & Iterate
Tasks are created as cron jobs. The first week is closely monitored. Data feeds the next iteration cycle.

Anti-Detection: The Hard Part

Generating LinkedIn content is relatively easy. Not getting caught is the real engineering challenge.

I built 3 layers:

Layer 1: NDI (Natural Dialogue Index)

Every engagement session receives a naturalness score from 1-10 based on:

Comment structure variation
Topic diversity (AI vs. non-AI)
Tool mention frequency
Rhetorical pattern repetition

If the session scores below 5.0, the system stops and recalibrates before continuing.

Layer 2: 7 Anti-Pattern Rules

Born from actual Day 1 mistakes:

Max 2 Claude/Anthropic mentions per 5 comments
Varied comment structures (never the same pattern twice consecutively)
At least 1 non-AI topic per session
Zero "evangelization" expressions
When someone agrees with us, just like — don't reply seeking another angle
Mandatory context research before commenting on posts with specific facts
No more than 15 comments or 30 likes per session

Layer 3: Epistemic Verification Gate

This was born from a real incident on Day 7. The system commented on a post citing a specific company case, and the AI-generated comment contained an inference that sounded factual but was wrong. The post author asked "what do you mean?" — exposing the gap.

Now, if a post cites a specific case, company, person, or event, the system MUST:

Read the full post (not just the preview)
Web search the case if it references specific facts
Never invent factual claims
Use opinion framing ("I wonder if...", "the pattern I see is...") when context is unclear

The Stack

ORCHESTRATION              EXECUTION                INFRASTRUCTURE
Claude Cowork              Chrome MCP               Google Cloud
  → 10 cron tasks           → LinkedIn DOM            → Cloud Run
  → Skills (.md)            → post publishing         → Vercel (site)
  → Sub-agents              → engagement sessions     → Sanity CMS
                            → screenshot capture

No intermediate tools. No Zapier, n8n, or Make. Claude IS the workflow — not a component in someone else's workflow.

Results (22 Days, Unfiltered)

Metric	Value
Followers	45 → 55 (+22%)
Engagement rate	3.0% (vs 2.21% baseline)
AI-written comments	75+
L1 proof events	13
Detection incidents	0
Engagement score	8.0/10 avg
Posts published	20/21 (95%)

The numbers aren't spectacular. They're real. And that's the point.

What Went Wrong

Chrome MCP instability: ~57% success rate in Week 1. The browser extension disconnected frequently during engagement sessions.
Day 1 pattern repetition: 3/5 comments had identical structure. Led to the 7 anti-pattern rules.
Day 7 epistemic failure: Wrong inference about a case cited by another professional. Led to the Verification Gate.
Engagement "evangelization": Comments like "I use Claude daily for this" sounded like brand promotion. Now capped at 1 per session max.

The Repo

GitHub: github.com/videomakingio-gif/claude-linkedin-automation

Install:

npx skills add videomakingio-gif/claude-linkedin-automation

Full writeup with all data: giovanniliguori.it/blog/claude-linkedin-automation-skill-open-source

The skill covers LinkedIn only. I have 21 total tasks covering blog, SEO, and operations, but only the 10 LinkedIn tasks are published — because publishing only what you can prove is more credible than promising everything.

Feedback and PRs welcome. Especially the uncomfortable kind.

DEV Community: Giovanni Liguori