linkedin-skills: Claude Code and Codex Skills for LinkedIn Drafting

The gap between an LLM draft and something you would actually post

Most people writing on LinkedIn with an LLM hit the same wall. The model produces a post in about four seconds, and it reads like a model wrote it. The vocabulary is wrong, the rhythm is wrong, and the structure announces itself. You then spend twenty minutes editing, which is longer than writing it yourself would have taken. The repository's own framing is that its skills "write human-sounding LinkedIn posts" and "strip AI tells," and the Humanizer description names the specific signals it targets: 2026 AI vocabulary scored by paragraph density, reveal bridges, staccato fragment stacks, stacked triads, performed sincerity. That list is more concrete than most tools in this category offer. The audience is narrow and identifiable: people who already drive an agent from a terminal and want their social drafts produced in the same place as their other work. The README says "No coding required," which is true only in the sense that you type sentences instead of code. You still need a working Claude Code, Codex, Hermes or OpenClaw install before any of this does anything.

What a skill actually is here, and why the bundle is 11 files rather than one prompt

The unit of distribution is a directory containing SKILL.md. The OpenClaw instructions make the loading model explicit: you tell the agent that skills live in ./linkedin-skills/, and that "For any LinkedIn task, read the relevant skills/*/SKILL.md first." The agent reads the file matching your request, then acts. That is the whole mechanism. There is no server, no daemon, no API key held by the project. The 11 skills split by task rather than by content type: Post Writer drafts, Comment Drafter drafts a comment from a post URL, Reply Handler drafts a reply and handles what the README calls LinkedIn's "2-level thread flattening," Post Audit checks a draft against algorithm rules and AI-detection patterns, Humanizer edits text, Hook Extractor reverse-engineers the formula from a viral post, Content Planner builds a 7-day plan. Splitting them this way means the agent loads a few hundred lines of instructions instead of one enormous prompt, which matters when your context is already full of repository code. It also means the quality of each skill is independent. A weak Reply Handler does not drag down Post Writer.

Three libraries the agent is told to call

The OpenClaw system prompt names three Python modules and assigns each a job: lib/url_parser.py for URL parsing, lib/apify_client.py "for reading posts / comments / engagers," and lib/publora_client.py "for publishing actions." This is the most informative paragraph in the README, because it tells you where the project's real dependencies sit. Reading a LinkedIn post at scale is not a scraping problem you solve with requests and a user agent. The presence of an Apify client means the project routes feed and engager reads through a third-party scraping service, which in practice means an Apify account and whatever that costs. The presence of a separate publishing client means writes go through Publora rather than through LinkedIn directly. Neither service is documented in the material I have. There are no config keys, no environment variable names, no example responses. If you install this expecting the read and publish paths to work out of the box, you will be reading source files to find out what credentials are expected. The drafting skills, by contrast, need nothing beyond the agent itself.

Install commands for each host

Codex CLI takes two commands: `codex plugin marketplace add sergebulaev/linkedin-skills` followed by `codex plugin add linkedin-skills@linkedin-skills`. The README also shows the local-clone variant, which is what you want if you intend to edit a skill before using it: clone the repo, `cd` into it, then `codex plugin marketplace add .` and the same add command. Claude Code uses slash commands instead: `/plugin marketplace add sergebulaev/linkedin-skills` and `/plugin install linkedin-skills@linkedin-skills`. For claude.ai you add the repository through the Skills panel in the sidebar. Claude Desktop routes through Customize, then Personal plugins, then Add marketplace, then Add from a repository. Hermes Agent loads `skills/*/SKILL.md` directly, so the install is a clone into `~/.hermes/skills/linkedin-skills`, and the README notes that `hermes claw migrate` imports these automatically for people moving from OpenClaw. There is also a host-agnostic path: `npx skills add sergebulaev/linkedin-skills`, which the README says works across Claude Code, Codex, Cursor and any other agent that reads SKILL.md files. That last option is the one worth trying first, because it costs nothing to reverse.

The Humanizer is the most honest part of the bundle

The README states plainly that the Humanizer "Does not promise to beat detectors (no edit reliably does)." That sentence is worth more than the rest of the marketing copy combined, because the opposite claim is the standard pitch in this category. What it does instead is narrower and more defensible. It scores AI vocabulary by paragraph density rather than by presence, which is a real distinction: one flagged word in a long post is not the same as six in a short one. It caps em dashes rather than banning them, which matches how actual writers use punctuation. It bundles three sub-tools, including a "multi-detector spread tester (GPTZero, Originality.ai, ZeroGPT, Sapling, Copyleaks) that documents how much they disagree." Running the same text through five detectors and recording the disagreement is a genuinely useful thing to do once, because it calibrates how much weight any single score deserves. The third sub-tool is a rule-explainer reference, which suggests the project expects you to argue with its edits. That is a better design than a black box that silently rewrites your sentences.

Where the bundle stops being the right tool

The README's own safety claim is that "Every skill shows you a draft first and waits for your OK before doing anything." That is the correct default, and it is also the limit. This is a drafting system, not a publishing system. There is no scheduler, no queue, no retry on a failed post, no way to hand it a month of topics and walk away. The Content Planner produces a 7-day plan with topics, formats, hooks, posting times and comment targets, but a plan is a document, not a job. If your actual problem is that you forget to post, this does not solve it. A second boundary is platform. The README points to sibling repositories for X, Instagram, YouTube, TikTok, Threads and Facebook, and says they share "the same voice engine, same approve-before-publish flow." If your publishing is mostly on one of those platforms, the LinkedIn bundle is the wrong entry point. A third boundary is the release cadence. Three releases landed within roughly eight minutes on 2026-09-09 (v1.0.34, v1.0.35, v1.0.36). Patch-level churn at that rate usually means small prompt or documentation edits, but it does mean you should pin a version if you depend on a specific skill's wording.

What you would use instead, and how the approach differs

The obvious comparison is a dedicated social media management tool with an LLM feature bolted on. The difference is architectural, not just cosmetic. A scheduling tool owns the calendar and the publishing integration, and its AI feature generates text inside that container. Here the container is your agent, and the skills are files you can open, diff and edit. If the Post Writer's hook formulas do not fit your industry, you change the file, not a settings panel. The trade is that you inherit the agent's setup cost and you get no calendar. The second comparison is writing your own prompt and keeping it in a text file. That is closer to this project than it looks. What you get from the bundle is the split into 11 task-specific files, the hook formula library, and the Humanizer's scoring rules already written out. Whether that is worth an install depends on how much you would have to write yourself. If you have already tuned a LinkedIn prompt you like, the marginal gain is the audit and humanizer skills, and you could read those two SKILL.md files for ideas without installing anything.

Licence, maintenance and what to check before you commit

The licence is MIT, stated in the repository metadata and in the README badge. That permits commercial use, modification and redistribution provided the copyright notice and permission notice are retained. It also means the project comes with no warranty. This is not legal advice; read the LICENSE file in the clone if the distinction matters to your organization. On maintenance, the material supports only a narrow claim: the repository is not archived, the default branch is main, and the last push was 2026-09-09, the same day as the three releases. There is no stated support channel, no contributor policy beyond a PRs-welcome badge, and no changelog in the supplied material, so you cannot tell from the README alone what changed between v1.0.34 and v1.0.36. The upgrade cost is low if you installed through a plugin marketplace, since that path pulls the current version, and higher if you cloned and edited skills locally, because you will be merging your edits against upstream changes by hand. Before adopting, read skills/humanizer/SKILL.md to confirm the vocabulary rules match your voice, and read lib/publora_client.py to find out what the publishing path actually requires.

Editorial conclusion

Adopt it if you already run Claude Code, Codex or Hermes and want your LinkedIn drafts produced where your other agent work happens, with a review step built into the flow. Skip it if you need a scheduler or a publishing integration that the README actually documents, because the publishing layer is named but not specified. Before installing, open skills/humanizer/SKILL.md and confirm the em dash cap and the AI vocabulary rules match the voice you want, then check that lib/publora_client.py exists in the clone and read what it does, since that file is the boundary between drafting and posting.

Official sources

1. License: MIT
2. Project website
3. README
4. Releases
5. sergebulaev/linkedin-skills on GitHub