AI automation
LinkedIn agent: posts, comments & DMs with Claude skills
In short
The LinkedIn Agent Skill is a free open-source pack (MIT licence) of eleven Claude skills: /li-post turns one idea into a finished post with three hook options, /li-comment writes comments on other people’s posts, /li-reply answers the thread under your own, /li-plan plans the week, /li-profile scores your profile out of 100, /li-carousel builds document posts, /li-repurpose turns one long piece into a week of posts, /li-dm and /li-inbox handle messages, /li-audit analyses your stats – and /li-human checks every draft for AI-typical patterns. The one rule: the skills write, you post. Nothing is published automatically, because that breaks LinkedIn’s User Agreement.
Free
- Open the repo on GitHub (github.com/avi691/linkedin-agent-skill) and copy the contents of skills/li-post/SKILL.md – plus hooks.json for posts, or rubric.json from li-profile for the profile check.
- Fill in the template templates/voice.md once: who you are, who you write for, how you sound, your positions and real numbers you are happy to use. Without this voice profile every draft sounds interchangeable.
- In a new chat (Claude or ChatGPT, free version) paste the skill text first, then your voice profile, then write your post idea underneath – the AI now works in “skill mode”.
- Pick one of the three hook options, check the finished text, replace placeholders such as {{your number}} with real values and paste the post into LinkedIn by hand.
Work according to the skill I pasted above (/li-post). My voice profile is below it. Write a LinkedIn post about this idea: [idea in 1–2 sentences]. First give me three hook options with their formula number, then the finished post: 900–1,300 characters, hook alone on line 1, short paragraphs, no links in the body, max 3 hashtags, no “Thoughts?” at the end.
Limits: Without installing, you lose the two Python scripts of the humanizer (/li-human) – the AI only checks for AI patterns by text rules. Skill text and voice profile must be pasted again in every new chat.
Premium / Automated
- Open Claude Code (type “claude” in a terminal or use the Code tab in the Claude desktop app) and paste this message: “https://github.com/avi691/linkedin-agent-skill – Install this skill, then confirm /li-post works.” Claude clones the repo and copies the eleven li-* folders to ~/.claude/skills/ – approve each step.
- Create your voice profile: paste three of your own posts and ask Claude to write ~/.claude/linkedin/voice.md from the template; then fill the “Proof” section with five real numbers (revenue, time saved, clients won, one costly mistake).
- Check your profile: feed /li-profile your headline, about text and last roles – you get a score out of 100 and rewrites in order of the biggest levers (headline and about first).
- Weekly routine (about 20 minutes a day): Monday /li-plan, every day /li-comment in batch for 5–10 posts, on post days “write Tuesday” → /li-post, /li-reply in the first hour after posting, Friday /li-inbox and /li-dm, monthly /li-audit with the LinkedIn analytics CSV.
/li-plan Here is what actually happened this week: [e.g. closed two client projects, lost one deal to a cheaper agency, rebuilt our onboarding]. Plan four posts (Proof, Opinion, Teach, Story) with angle and hook number, posting times for my B2B audience in the [Europe/London] timezone, and an engagement list of ten people (5 reach, 3 peers, 2 buyers).
Added value: All eleven skills as slash commands, humanizer scripts run locally (five checks: burstiness, specificity, slop density, fingerprint, voice – PASS from 70 points), weekly plan and post log are saved in ~/.claude/linkedin/, updates via the /plugin marketplace. No API key, no LinkedIn login needed.
Frequently asked questions
Does the agent post to LinkedIn automatically?
No, deliberately not. Automated posting via browser bots or unapproved tools breaks LinkedIn’s User Agreement and can get your account restricted. Every skill ends with a copy-ready block of text that you read, approve and paste yourself.
Do I need coding skills or an API key?
No. For the easiest path you paste one message into Claude Code – it installs everything itself and asks before each step. There is no API key, no LinkedIn login and no third-party account. The two humanizer scripts only need Python 3, which Macs usually already have.
Won’t the posts sound like AI?
That depends on your voice profile (voice.md): who you are, how you write, your positions and real numbers. Without it, drafts sound interchangeable. In addition, the humanizer removes invisible characters, em dashes and 113 stock AI phrases and scores every draft on five checks – but it explicitly does not guarantee anything against AI detectors.
Does the AI invent numbers or client examples?
No. The skills never invent metrics, clients or results. If a value is missing, the draft comes back with the placeholder {{your number}} and a flag – you fill in the real figure.