AI Guides › Workbench
By Nigel Guy · 7 min read
Most people who install a LinkedIn skill pack ask for a post, paste the result and publish it. It feels efficient because the draft arrives fast and reads fine. The cost shows up later: the same stock vocabulary, the same punchy fragments, the same reveal in the last line, and readers who have seen all of it before.
The rule: let the pack draft, but run every piece through the humaniser and your own voice before it goes anywhere near LinkedIn, and never let any skill post, comment or message on your behalf.
The pack is an open-source repository, sergebulaev/linkedin-skills, published under an MIT licence. Its README describes skills for Claude Code and Codex that draft LinkedIn content from your terminal, each showing you a draft and waiting for your approval. I read the README on 2026-10-04 and have not audited the code line by line.
One thing to know first: you may have heard "eleven skills", but the README I read lists twelve. It adds a post audit and an interviewer (a story bank) to the eleven you will see in many skill listings. Counts change between releases, so check the repository's own list.
| Skill | What it does | Cost | Best for | Catch |
|---|---|---|---|---|
| Post writer | Drafts posts using a set of hook formulas | Free (MIT) | Getting a first draft down | Formulas produce familiar shapes |
| Comment drafter | Drafts a comment on a post you give it | Free | Thoughtful replies you then edit | Generic comments are worse than none |
| Reply handler | Drafts replies in comment threads | Free | Keeping up with a busy thread | You must read the thread yourself |
| Post audit | Checks a draft against algorithm rules and AI patterns | Free | A last check | "Algorithm rules" are the author's claims, not LinkedIn's |
| Humaniser | Strips common AI tells from text | Free | Every draft | Does not guarantee anything (see below) |
| Hook extractor | Breaks down why a viral post is built the way it is | Free | Learning structure | Copying structure is not the same as having something to say |
| Content planner | Builds a 7-day posting plan | Free | Getting unstuck | A plan is not a reason to post daily |
| Engagement monitor | Tracks comment threads and who is engaging | Free skill; reading posts can use an optional Apify token | Seeing who replies | Data collection has terms issues (see Guardrails) |
| Profile optimiser | Rewrites your profile for conversion | Free | Headline and About section | Check every claim is true |
| Employee advocacy | Plans a team-wide programme | Free | Small companies | Nobody should be pressured to post |
| Repurposer | Adapts other content for LinkedIn | Free | Turning a blog post into a post | Cut and rewrite, do not paste |
| Interviewer | Builds a personal story bank | Free | Material only you have | Needs real, honest answers |
The skills themselves are free. The README lists optional extras: an Apify token for reading posts and comments (Apify's monthly credit is $5, about £4 at time of writing), a Publora key for publishing (a free allowance of 15 posts a month is quoted), and an image-generation token. All of it reportedly works without keys by copy and paste. Check current prices and the £ figure on each vendor's page before signing up.
In a Claude Code session, add the repository as a marketplace, then install from it:
/plugin marketplace add sergebulaev/linkedin-skills
/plugin install <plugin-name>@<marketplace-name>
The marketplace and plugin names are printed when the first command succeeds, or visible on the Discover tab when you run /plugin. Choose an install scope when asked: user scope for every project on your machine, or local scope for one repository. Then type / and look for the skills, which Claude Code namespaces as /<plugin>:<skill>.
The README also describes adding it through the Plugins area of Claude Desktop and claude.ai (Customize, then Plugins, then Add marketplace, then paste the repository name). I could not test those routes, and the menu labels move often, so follow what your screen shows.
Two housekeeping points from Anthropic's plugin documentation. Third-party marketplaces do not auto-update by default, so update the plugin yourself when you want a new version. And a plugin can bundle hooks and MCP servers, so read the details pane before you confirm the install.
The trap is treating the humaniser as optional polish, or as a disguise. It works on the surface of the text: the README says it regulates em-dash density (roughly one per 100 words), removes stock AI vocabulary such as "leverage" and "unlock", and flags stacked fragments and reveal-style bridges. It can also compare several AI-detection services, which are known to disagree with each other. The README states it does not promise to beat detectors.
So there are two ways to get it wrong. One is believing a passed check means your post cannot be spotted. The other is believing a spotted post is the main risk. The real risk is a feed full of text that sounds like everyone else's. The humaniser removes the tells; only your own facts and opinions supply the substance.
You are a careful editor helping me finish a LinkedIn post in my own voice.
Context: I will paste a draft, then two samples of my real writing, then the facts I can vouch for.
Goal: a post I would be comfortable defending in a meeting, which sounds like my samples, not like a template.
Steps:
1. Read my samples and note sentence length, vocabulary and how I open and close.
2. Compare the draft. List every sentence that contains a claim, figure or example not in my facts list, and ask me about each one rather than keeping it.
3. Rewrite the draft to match my voice. Remove stock phrases and stacked short fragments. Keep the structure only where it helps.
4. Return the revised post, then a short list of what you changed and what you could not verify.
Constraints: British English. No invented statistics, quotes, client names or results. If a sample or fact is missing, ask for it before writing. Before answering, check that every factual claim traces to my facts list.
[DRAFT]
[SAMPLE_1]
[SAMPLE_2]
[FACTS_I_CAN_VOUCH_FOR]
Fill in the four bracketed inputs; the more real material in the samples and facts, the less the model has to guess.