v1.0.0

li-audit

jakeschincariol@jakeschincariol⭐ 3.5k stars· last commit 2mo ago· 7 open issues

>-

Choose how to use this skill

You do not need every option. Choose the path your AI client supports. The stable page stays the same; versioned files are immutable.

1. Native installer

This listing has no registered native installer command. Use the complete package or source fallback below, depending on what your client supports.

Do not guess an installer command or replace an existing version without reviewing the diff.

2. Complete package recommended

Download the ZIP when available. It includes SKILL.md plus the references, security notes and version metadata.

No complete ProSkills package is published for this listing yet.

3. Prompt-only

Copy the prompt above when the agent can read the stable page or when you want to adopt the workflow without installing a skill.

Need only the instruction file?

Download SKILL.md only if your client requires a single file. The complete ZIP is safer for a full installation because it preserves the references and release context.

No path installs or executes anything by itself. Your agent still needs access to the project files. Before updating, compare the installed version and review the diff.

7.2/10
Verified
Oct 12, 2026

// RATINGS

⭐GitHub Stars
⭐⭐⭐⭐⭐ 3.5k on GitHubGitHub ↗

Very popular

🟢ProSkills ScoreAI Verified
7.2/10
📍

Not yet listed on ClawHub or SkillsMP

// README

The LinkedIn agent skill

Eleven Claude skills that run a LinkedIn account. Free, MIT, no signup, no API key, nothing to connect.

One of them writes your posts off 21 hook formulas. One comments on other people's posts. One handles the replies under yours. One scores your profile out of 100 and rewrites what lost points. One plans the week: what to post, when, and who to engage with.

And one is the humanizer, which is the reason the rest are usable. It strips the em dashes, the slop vocabulary and the invisible watermark characters out of a draft, then scores what is left against a five-check detection panel before you ever see it.

Nothing gets posted until you say yes. These skills write. You post.

Install

Paste this into Claude:

https://github.com/Jakeschincariol/linkedin-agent-skill

Install this skill, then confirm /li-post works.

Or do it yourself, in Claude Code:

git clone https://github.com/Jakeschincariol/linkedin-agent-skill.git
cp -r linkedin-agent-skill/skills/li-* ~/.claude/skills/

Or as a plugin:

/plugin marketplace add Jakeschincariol/linkedin-agent-skill
/plugin install linkedin-agent

Project-local instead of global: copy the same folders into your repo's .claude/skills/. No Claude Code at all? Paste any single SKILL.md at the top of a chat and it runs as a mode - you lose the two Python tools, which is most of the point of /li-human, but the rest works.

Then spend ten minutes on templates/voice.md. Copy it to ~/.claude/linkedin/voice.md and fill it in, or paste three of your own posts into Claude and say "write my voice.md from these". Every skill reads that file. Skip it and everything comes out sounding like everyone else.

The eleven

commandwhat it does
/li-postOne idea into a post. Three hook options from 21 formulas, one full draft, humanized before you see it.
/li-commentComments on other people's posts. Nine types, picked by what the post actually is. Never "Great post!".
/li-replyThe thread under your own post. Sorts every comment into lead / substance / peer / support / noise, then writes in that order.
/li-profileScores your profile against a 12-part rubric out of 100, then rewrites in fix-first order.
/li-planThe week. What to post, when to post it, and the 10 people to engage with. Writes ~/.claude/linkedin/plan.md.
/li-humanThe humanizer. Two scripts that actually run. See below.
/li-carouselDocument posts. Slide-by-slide copy, the cover that earns the swipe, and the PDF to upload.
/li-repurposeOne video, newsletter or transcript into a week of posts that each stand alone.
/li-dmThe 200-character invite note, the first message, and the two follow-ups. Two.
/li-inboxTriages the inbox into lead / recruiter / peer / ask / spam, and tells you which tell gave the sequence away.
/li-auditPost-mortem on what you have already published. Ranks by engagement rate and reach multiple, not impressions.

The humanizer

/li-human ships two Python scripts with no dependencies. They run on your machine, on your text, and nothing is uploaded.

python3 humanize.py draft.txt --report      # clean it, show every change
python3 detect.py draft.txt                  # score it, five checks
python3 detect.py before.txt after.txt       # prove the delta

What comes out automatically:

  • Invisible characters. Zero-width spaces and joiners, word joiners, soft hyphens, byte-order marks, Unicode tag characters, non-breaking and narrow spaces. Your keyboard does not make these. They survive copy-paste and they are invisible in every editor you own.
  • Typography. Em dash to comma, en dash to hyphen, curly quotes to straight, ellipsis to three dots.
  • The lexicon. 113 stock words and phrases with plain-English replacements - delve, leverage, robust, seamless, crucial, testament to, "in today's fast-paced world", "let that sink in" - with capitalisation preserved and URLs untouched. It lives in slop.json and it is meant to be edited.

What gets flagged instead of fixed: "It's not just X, it's Y", rule-of-three triads, one-word rhetorical questions, hashtag walls, reflex engagement bait, uniform sentence length. Changing the shape of a sentence needs judgement, so those are handed back for a rewrite rather than mangled by a regex.

The five checks, scored 0-100, higher is more human:

checkwhat it measures
BURSTINESSsentence-length variation. Models write even.
SPECIFICITYnumbers, names and concrete markers per 100 words
SLOP DENSITYlexicon hits per 100 words
FINGERPRINTinvisible characters, em dashes, curly quotes per 1,000
VOICEcontractions, person, structural tells

The verdict weights the mean at 60% and the weakest single check at 40%, because a detector only needs one signal to fire.

Run against a deliberately terrible draft:

  BURSTINESS    ##################......  73.0
  SPECIFICITY   ######################## 100.0
  SLOP DENSITY  ........................   0.0    19 stock terms, 24.1 per 100 words
  FINGERPRINT   ........................   0.0    1 invisible, 1 em dash, 3 curly quote
  VOICE         ########................  33.3    3 structural tells
  ------------------------------------------------------------
  HUMAN SCORE   ######..................  24.8   FLAGGED

After humanize.py, with the flagged structures still unrewritten:

  HUMAN SCORE   #################.......  69.7   REVIEW    (+44.9)

The last stretch to PASS is the part the script deliberately leaves to you.

The fine print, which is the honest part

These skills do not post to LinkedIn, and they should not. There is no official API for posting to a personal profile without an approved partner app, and automating the site with a browser or a third-party tool violates LinkedIn's User Agreement and gets accounts restricted. So every skill here ends the same way: a copy-ready block, and you paste it. That is not a limitation bolted on afterwards, it is the design. It is also why the approval gate is real rather than a setting.

The five checks are local heuristics, not detector APIs. They are modelled on the signals public detectors key on, and they run entirely on your machine. They are not GPTZero, Originality, Copyleaks, Winston or Turnitin, they do not call those services, and they cannot promise those verdicts. Fixing what they measure tends to move those numbers, because they are measuring the same underlying things. That is the whole claim. Nobody can honestly sell you "undetectable", and anybody who does is selling you something.

The invisible-character pass is real and it is narrow. It removes the zero-width and format characters that end up in generated text and survive a copy-paste. That is a genuine, checkable fingerprint. It is not a claim about defeating a cryptographic watermarking scheme, and this repo does not make one.

Nothing here fabricates. No invented metrics, clients or outcomes go under your name. If a draft needs a number you have not given, it comes back with {{your number}} in it and a flag, every time.

Files

skills/li-post/hooks.json        21 hook formulas: template, example, what it is for, how it gets ruined
skills/li-human/slop.json        the lexicon: 113 terms, 17 invisible classes, 11 structural tells
skills/li-human/humanize.py      the three cleaning passes
skills/li-human/detect.py        the five-check panel
skills/li-profile/rubric.json    the 100-point profile score
templates/voice.md               your voice profile. Fill this in first.

Credit

Made by Jake Schincariol, opusjake.ai. The full write-up is at [opusjake.ai/r

// HOW IT'S BUILT

// FILES

2 files shipped with this skill. Inline text is the catalog copy. The upstream link is the source repository.

  • skills/li-audit/SKILL.md3.3 KBupstream ↗
    ---
    name: li-audit
    description: >-
      Post-mortem on what the user has already published - which posts actually
      worked, why, and what to stop doing. Use when the user pastes their LinkedIn
      analytics or past posts and asks "what's working", "why did this flop", "read
      my analytics", "audit my content", or wants to know what to double down on.
    ---
    
    # li-audit
    
    The only honest source of what works for an account is that account. Every
    rule in every LinkedIn guide, including the ones in this pack, is a prior. The
    user's own last 30 posts are the evidence.
    
    ## Input
    
    Ask for whichever the user has:
    
    - The post analytics export (LinkedIn: Analytics -> Content -> Export). CSV.
    - Or a screenshot per post with impressions, reactions, comments, reposts.
    - Or just the posts and their reaction counts, which is enough for a first
      pass.
    
    Also read `~/.claude/linkedin/log.md` if it exists, since it records which
    hook formula each post used.
    
    ## What to actually measure
    
    Raw impressions are the least useful number on the page, because they are
    mostly a function of how many people already follow the user. Compute these
    instead, and show the working:
    
    | metric | how | what it tells you |
    | --- | --- | --- |
    | **Engagement rate** | (reactions + comments + reposts) / impressions | whether the post earned its reach |
    | **Comment ratio** | comments / reactions | whether it started something or just got a nod |
    | **Reach multiple** | impressions / follower count | whether it travelled past the existing audience |
    | **Save/send rate** | if available | the strongest single predictor of future reach |
    
    Rank by engagement rate and reach multiple, not impressions. A post with 900
    impressions and 40 comments beat the one with 12,000 impressions and 6.
    
    ## Then find the pattern
    
    With the top 5 and bottom 5 side by side, look for what actually separates
    them, and be willing to conclude something the user will not like:
    
    - Hook formula. Which numbers from `hooks.json` are in the top 5?
    - Format. Text, document, image, video.
    - Length.
    - Theme.
    - Day and time - check this **last**, and only if the other four show nothing.
      It is almost never the cause, and it is where people want it to be.
    - First-hour comments. Posts the user replied to inside an hour versus not.
    
    State the finding as a claim with the evidence attached, and say how confident
    it is. With 30 posts you can see a pattern; with 6 you cannot, and you should
    say that instead of inventing one.
    
    ## Output
    
    ```
    AUDIT  ·  31 posts  ·  Jun 12 - Sep 5
    
    TOP 5 BY ENGAGEMENT RATE
      8.1%  #3  Mistake      "$18,000 is what no contract cost me"      1,940 imp
      6.4%  #20 Walk-Away    "I fired my highest-paying client"         2,210 imp
      ...
    
    BOTTOM 5
      0.4%  #5  List         "7 tools every founder needs"             11,400 imp
      ...
    
    WHAT THE DATA SAYS
    1. Posts where you were the one who looked bad: mean 6.2% vs 1.1% for
       everything else. n=6. This is your strongest signal and it is not close.
    2. Tool listicles get impressions and nothing else. High reach, no comments,
       no leads. Three of your bottom five.
    3. Day of week shows nothing. Your Tuesday mean and your Friday mean are
       inside the noise. Stop optimising it.
    
    STOP: listicles about tools.
    DO MORE: the ones with a cost you paid, and a number.
    ```
    
    Then hand the conclusions to `/li-plan` so next week's plan is built on the
    user's own evidence rather than on defaults.
    
  • README.mdupstream ↗

    Contents are not inlined in the catalog. Open the upstream file.

Plain-text copy, no login: /skill-bundles/jakeschincariol-linkedin-agent-skill-skills-li-audit.md

// REPO STATS

3.5k stars
7 open issues
Last commit: 2mo ago