Claude LinkedIn Automation

Autonomous LinkedIn management, validated in production.
27+ days. 10 tasks. Zero detection. 3.9% engagement rate.

Install  •  How It Works  •  Results  •  Anti-Detection  •  Compatibility  •  Contributing

Legal Disclaimer — This skill documents an autonomous LinkedIn management system. Automated interactions may violate LinkedIn's User Agreement. Use at your own risk. The authors assume no liability for account restrictions or bans. Published for educational and research purposes.

What Is This?

A custom skill for Claude that turns your AI assistant into a full-stack LinkedIn manager. It posts daily, engages with your network, triages DMs, audits itself for detection risk, and reports weekly — all autonomously.

Every rule is extracted from 27+ days of real production data on a live Italian profile. Not theory. Not best guesses. Empirical evidence from daily audits, scored engagement sessions, and documented incidents that shaped the system.

Works in any language. The wizard was battle-tested in Italian, but the system is language-agnostic. Phase 1 captures your identity, voice, and vocabulary in whatever language you operate in — Claude generates all content in your language. The architecture (pillar calendar, anti-detection rules, NDI scoring, task scheduling) is universal.

The 5-Phase Wizard

Type /linkedin and Claude walks you through setup:

Phase 1  IDENTITY        15 questions to define your voice, vocabulary, red flags
Phase 2  STRATEGY        7-day pillar calendar, post format, humanization rules
Phase 3  ENGAGEMENT      Commenting rules, anti-detection, epistemic verification
Phase 4  TASK PLAN       Review every task before anything gets automated
Phase 5  CREATE & RUN    Deploy tasks, monitor, iterate weekly

Nothing is automated until you explicitly approve. Phase 4 is a hard gate — Claude will not proceed without your "approved."

Install

git clone https://github.com/videomakingio-gif/claude-linkedin-automation.git
cd claude-linkedin-automation
chmod +x install.sh && ./install.sh

The interactive installer walks you through 3 choices:

Quick install (skip the wizard):

./install.sh --global     # All projects, Claude Code
./install.sh --project    # Current project only
./install.sh --uninstall  # Remove everything, all IDEs

Update: git pull — Claude Code uses a symlink, so the skill stays in sync. Cursor/Windsurf use file copies — re-run the installer after pulling.

After installing, type /linkedin in any Claude session to start.

How It Works

Architecture

claude-linkedin-automation/
├── SKILL.md                              # The skill itself (5-phase wizard)
├── HUMAN-VOICE-LAYER.md                  # Anti-detection Level 2: structural naturalness
├── install.sh                            # Interactive installer
├── modules/
│   └── linkedin.md                       # Full module config (560 lines)
├── references/
│   ├── tov-framework.md                  # Voice: 10 rhetorical patterns, vocabulary, registers
│   ├── anti-detection-playbook.md        # 7 rules (L1) + Level 2 structural tells, NDI formula
│   ├── content-templates.md              # Day-by-day templates with worked examples
│   ├── epistemic-verification.md         # 7-checkpoint fact verification gate
│   └── task-catalog.md                   # Full prompt templates for all 10 tasks
├── examples/
│   ├── weekly-plan.md                    # Real Week 3 content plan
│   └── engagement-session.md             # Scored session with 5 comments
├── assets/                               # Growth charts and dashboard
├── CHANGELOG.md
├── CONTRIBUTING.md
└── LICENSE

The 10 Tasks

Minimum viable setup: Tasks 1 + 2 + 9. Three tasks, fully autonomous.

Results

27+ Days of Production (April 6, 2026)

Professionals replied by name, sent multi-message DMs, mentioned the profile in their own posts, and sent connection requests — all without suspecting automation.

Key milestone: On April 4, 2026, the system completed a full attribution cycle: LinkedIn post → site visit → purchase of a digital product. The funnel worked without any manual intervention.

Growth Charts

Interactive Dashboard (HTML) — hover tooltips with daily data

Anti-Detection

The system uses a two-level anti-detection architecture.

Level 1: Behavioral Rules (7 rules, empirically validated)

Level 2: Structural Naturalness (Human Voice Layer)

Level 1 prevents algorithmic flags. Level 2 addresses a subtler problem: pattern recognition by expert human readers. Even with perfect vocabulary and timing, certain structural tells betray AI authorship to the professionals who matter most.

The 7 structural tells that L1 doesn't cover:

6 alternative post structures are defined in HUMAN-VOICE-LAYER.md: Stream of Consciousness, Question Without Answer, Start From the Middle, Broken List, Micro-post, Response to Something.

Pre-publication checklist (5/7 required to publish):

• Different structure from yesterday and the day before?
• No perfect parallelism in lists? (at least 1 asymmetric element)
• At least 1 genuine flow break? (not an inserted marker, a real interruption)
• Numbers are not all round? (not 85→9, but 85→11 or "something like 80-90 mins")
• Case study has dirty details? (vague memory + specific detail)
• Emotional arc is not always positive? (at least 1 unresolved post/week)
• Could this post have been written by a human in 5 minutes?

Full methodology: HUMAN-VOICE-LAYER.md

Non-Detection Index (NDI)

NDI = (L1 × 2 + L2 × 1) / (L1 + L2 + L3) × 10

• L1 (weight 2): Named replies, multi-message DMs, public mentions
• L2 (weight 1): Genuine questions, connection requests
• L3 (weight 0): Generic likes, one-word replies

NDI > 5.0 = healthy. < 3.0 = investigate. < 4.0 two weeks = pause 48h and audit.

Epistemic Verification Gate

Before publishing any factual claim, run 7 checkpoints:

1. Fact vs. Inference — label it correctly
2. Uncertainty Markers — verified / observed / inferred / speculative
3. Source Attribution — name it or don't claim it
4. Temporal Coherence — when did this happen?
5. Case-Specific Claims — verify in 30s or rephrase as question
6. Self-Assessment Bias — measured vs estimated vs projected
7. Absence-as-Proof — "I haven't found" ≠ "it doesn't exist"

7/7 pass = publish. 5-6/7 = fix and publish. <5/7 = rewrite.

Full methodology: references/anti-detection-playbook.md

Compatibility

Recommended Setup

Reference Files

Niche Adaptation

The skill was built in the AI/B2B automation niche, but the architecture is domain-agnostic. The 7-day pillar calendar adapts to any niche — you keep the emotional structure, change the content domain:

The anti-detection rules, NDI scoring, epistemic verification gate, and task scheduling work identically across all niches. Only the content domain and vocabulary change — and the wizard captures those in Phase 1.

FAQ

How long does setup take?
4-6 hours total. Identity definition is 2-3 hours (the hardest part). First week of content 2-3 hours. Task creation 15 minutes.

What's the minimum viable setup?
Tasks 1 (daily-post) + 2 (daily-engagement) + 9 (weekly-report). Three tasks, fully autonomous.

Can I post more than once per day?
Don't. LinkedIn penalizes same-day multiple posts.

How do I know if comments are natural?
Target 8.0+/10 on the scoring rubric. Below 7.0 = adjust rules. See anti-detection-playbook.md.

What if a task fails silently?
Every task writes a log. Check report/ daily. The experiment-audit task (daily 15:00) catches most silent failures.

Does this only work in Italian?
No. The wizard and examples are in Italian (the production language), but the system works in any language. Phase 1 captures your voice, vocabulary, and audience in your language — Claude generates everything accordingly.

Does this only work for the AI/automation niche?
No. The architecture (pillar calendar, anti-detection, NDI, verification gate) is niche-agnostic. See Niche Adaptation for examples.

What is the Human Voice Layer?
It's a Level 2 anti-detection framework added after 22 days of operation. Level 1 prevents algorithmic detection. Level 2 addresses structural patterns that reveal AI authorship to expert human readers — even when vocabulary and timing are correct. See HUMAN-VOICE-LAYER.md.

Who Built This

Giovanni Liguori — AI Automation Architect

I transform manual processes into automated ecosystems for Italian SMBs and freelancers using Claude + Python + Google Cloud.

giovanniliguori.it  •  LinkedIn  •  Case Study

The Experiment

Can a well-instructed LLM manage a professional LinkedIn profile without being identified as non-human?

After 27+ days of daily operation:

• Zero detection incidents
• 13+ L1 proof events (named conversations with professionals)
• 8.0/10 average engagement quality
• 3.9% average engagement rate
• 5.0+ NDI (Non-Detection Index) consistently
• 1 product sale attributed directly to the LinkedIn funnel (April 4, 2026)

The system works because it treats identity and anti-detection as the same thing. A profile with a clear, consistent, humanized voice is inherently less likely to be flagged. It's also more likely to convert.

The Level 2 (Human Voice Layer) extends this principle: structural naturalness — varied post formats, asymmetric lists, dirty case study details, unresolved emotional arcs — builds the kind of trust that drives DMs, connection requests, and ultimately sales.

License

MIT License — Copyright (c) 2026 Giovanni Liguori

Contributing

See CONTRIBUTING.md. The methodology improves with more data points.

_{Built with Claude. Validated in production. Open source.}