Do I need Claude Code specifically, or will this work with another agent?

The repo describes the skills as being for Claude Code and Codex, so it is not Claude only. The voice file, the de-AI pass and the approval gate are plain instructions, so you can paste them into whatever tool you already use. Only the folder based install is specific to agents that load skills from a directory.

How many skills are in the bundle?

It depends on the source. The author's own write-up says the project started as ten skills, while his GitHub profile and a third party skills index both list 11. Open the repo and count the folders before you rely on a specific one being present.

Will this stop my posts sounding like AI wrote them?

Only if you do the voice file properly. A de-AI or humanizer step removes obvious tells such as stock phrases and padded transitions, but it cannot invent an opinion you have not written down. The rules, the banned words and the proof list are what make the output yours.

Set up a LinkedIn writing workflow with Claude skills in about three hours

A public bundle of LinkedIn skills for Claude Code will not fix your company page. What it can do is force you to write your house style down once, so a machine applies the same rules twice instead of guessing every morning.

The repo is sergebulaev/linkedin-skills. I read the repo page, the author's write-up and a third party index of it. I have not run the bundle end to end, so every minute figure below is my estimate, not a measurement.

What is actually in the bundle

The repo presents itself as LinkedIn marketing skills for Claude Code and Codex. The description says the skills write posts, comments and replies in your voice, strip AI tells, and wait for approval before anything gets published (repo).

The author's own write-up says the project started as ten open source skills covering the full loop: plan, write, de-AI, audit, publish, comment, reply, analyze, optimize and scale (bulaev.net).

The count depends on who is counting

The author's GitHub profile summarizes the project as 11 Claude Code and Codex skills for human sounding posts, comments, feed analysis and publishing cadence (profile). A third party skills index also lists 11, with named entries such as linkedin-marketing, linkedin-profile-optimizer and linkedin-humanizer (skills.sh).

So: ten in the author's post, 11 on the profile and the index. The bundle is young and the grouping moves. Treat any article that states a hard count, including the headline that sent me looking, as approximate.

What this means for a small business

One sentence: this is a content ops kit, not a ghostwriter replacement, and the part worth your time is the workflow around the draft, not the draft.

Any model will hand you a LinkedIn post. That was never the bottleneck. The bottleneck in a founder led team is that posts go out sounding like every other post, and nobody reviews them against anything written down. The bundle's answer is three steps most people skip: a de-AI pass, an audit, and an approval gate before publish. That structure is worth copying even if you never install a single skill.

I run a content engine that ships 40+ posts a month across 14 platforms. The drafting was never the slow part. Writing down the rules and the proof list was.

The setup, step by step

Time budgets are my estimates from reading the sources, not a stopwatch.

Read the repo and every SKILL.md: about 20 minutes. A skill is instructions a model will follow with your account attached. Read each one before you run it, the same way you would read a script someone handed you.
Install: about 15 minutes. Claude Code loads skills from a skills directory, ~/.claude/skills for user level skills or .claude/skills inside a project. Clone the repo and put the skill folders there. If the README specifies a different path or an install command, the README wins over this paragraph.
Write your voice and rules file: about 45 minutes. This is the step that decides whether your posts stop looking like everyone else's. Skeleton below.
First draft through the writer skill: about 30 minutes. Give it a real topic you already know the answer to, so you can judge the output.
De-AI pass and audit: about 25 minutes. Run the humanizer and audit steps, then mark by hand every sentence you would not have said out loud. That list becomes rule additions.
Approval checklist: about 20 minutes. Write down who says yes and what they check. Nothing posts without it.

Total: about two and a half hours, inside the three in the title, with room for the install going sideways.

The voice and rules file skeleton

Six blocks. Write it once, paste it into every skill that drafts.

Reader. One named role, not "businesses". The estimator, the office manager, the ops lead.
Three sentences you would never write. Concrete ones, copied from posts you dislike.
Banned words. Ten of them, the ones your industry over-uses.
Proof list. The only numbers a post may state, and where each one came from. If it is not on this list it does not go in a post.
Format rules. Opening line, paragraph length, no wall of hashtags, where the link goes.
The gate. Who approves, and the two things they check before yes.

Skip the proof list and you will eventually publish an invented client story under your company name.

What the skill does and what stays with you

Step	Let the skill do it	You decide
Topic selection	Suggest from the feed	Which one is on strategy
First draft	Yes	Nothing
De-AI pass	Yes	Whether it still sounds like you
Numbers and claims	No	Every figure, against the proof list
Client or customer stories	No	All of it
Publish	Only after a human yes	The yes

Limitations

I have not run this bundle end to end. The time budget is an estimate and the install step could easily take longer on a first try.
The skill count is unresolved across the sources. If a specific named skill matters to you, open the repo and check it is still there.
Skills are instructions, not a control system. Nothing in a prompt prevents a model inventing a number. Your proof list and your reviewer do.
Anything that publishes on your behalf needs credentials. Check what each skill touches before you hand over an account.
Voice transfer is the weak point. My guess: a thin voice file produces posts that read like the average of LinkedIn, and no humanizer step fixes that.

My take

The useful idea here is not "AI writes LinkedIn posts". It is that content production has stages, and the stages after the draft are where the quality lives. You can get most of the benefit in an afternoon with a written voice file and an approval gate, with or without this repo.

If you want the repo, clone it, read the SKILL.md files, and spend your time on the voice file. If you want the idea, write the six blocks today and point whatever tool you already use at them.

Next step: pick one post you published last month, and write down the three rules it broke. That is the start of your file. Other write-ups on building this kind of pipeline are on the blog.