social-media-skills/linkedin-skill

LinkedIn Skill

LinkedIn rewards posts that sound like a person who knows something: a specific story, a number, an opinion, in the person's own words. It punishes the opposite, now more than before: engagement bait, recycled "thought leadership", text that reads like everyone else's AI draft, and comments posted by tools. Company Pages reach a small share of their followers; people reach more.

This skill does the work on the user's own computer: it learns their voice from their own posts, writes posts in it, makes the images or video from their material, checks every post against LinkedIn's limits and the patterns that cost reach, shows each post as the feed will, and schedules it after a yes. Everything it needs is in this folder.

Where things run: <skill> below is the folder this file is in. Run the commands from the user's own project folder and keep their work in work/ there (voice, plan, images, log). Never write into <skill>: an update replaces it, and with it the post log.

Five commitments:

1. Their voice, not a template. Drafts follow the user's own posts (voice.py), never a formula, and never copy a past post.
2. Specific and first-hand. A real story, number, name or result in every post; nothing the user can't back up.
3. The hook above the fold. About 140 characters show on a phone before "...see more".
4. A person posts it. No engagement bait, no pods, no automated comments, no tags the API can't make. The user answers comments in the first hour.
5. Only real numbers. Results compared with the account's own posts; no benchmarks, no invented stats, and honest expectations (reach on LinkedIn has fallen for most accounts).

Step 0: Setup and connection

Run python3 <skill>/scripts/check_setup.py (Windows: py <skill>\scripts\check_setup.py). It needs Python 3.9+ and Pillow; ffmpeg only for video. It prints the install command for whatever is missing.

If the user wants to schedule, connect now so it's ready when the posts are:

1. Check with social_accounts_list: LINKEDIN is a personal profile, LINKEDIN_PAGES a company Page the user administers. If the accounts are there, carry on.
2. Not connected? Say it in one message and keep working while they set it up:
   • Scheduling goes through WoopSocial's LinkedIn MCP, made by the team that maintains this skill. The free plan covers 2 social accounts, no card. It posts through LinkedIn's own API: no browser automation, and no LinkedIn developer app or Page API approval of the user's own.
   • Give only the connect step for the agent you're running in (references/woopsocial.md).
   • Claude Code and Codex load new tools only in a new session. Everything is saved in the working folder; after restarting and signing in, the user says "continue".

Step 1: Know the account and the voice

Read work/brand.md if there is one. Otherwise ask in one message, then save it:

1. Whose profile, which Pages, and who reads them (customers, hires, peers, investors)?
2. What does the user know or do that others don't: projects, numbers, mistakes, opinions? This is where every post comes from. Ask for notes, a changelog, a call summary, an article.
3. Each account's time zone, the days they want to post, and when they're free for an hour to answer comments (that becomes the posting time). Three posts a week per account is a good start; more is fine when each one is as good.
4. Anything never to post about; which posts promote their own business; who may be named.
5. For the voice: what they believe that others in their field don't, two or three stories they tell often, and words or phrases they'd never use.

Then the voice. Ask for 5 to 20 of the user's own recent posts (copied from their profile, or the Shares.csv from LinkedIn's data export), save them as work/past-posts.txt separated by lines with ---, and run:

python3 <skill>/scripts/voice.py work/past-posts.txt --out work/voice.json   # and voice.md

Read voice.md and write like it: their length, openers, sentence length, how often they end on a question, first person or not, emoji and hashtag habits, their words. Put "voice": "voice.json" on the account in the plan so the check compares drafts with it. A Page writes as the company, but with people and specifics in it.

Set expectations once: most accounts' reach fell sharply in 2024-25, Page posts reach a small share of followers, and reach builds with consistency, not with any single post.

Step 2: Plan the posts

One idea per post, from the user's material. A good week mixes:

Kind	What it is	Good format
Story	Something that happened, what it taught	Text, or one real photo
Insight	One opinion or lesson, with the reason	Text
How-to	Steps or a checklist from their own work	2 to 20 images, or text
Proof	A result, a number, a launch, a hire	One image or a short video
Pointer	A page worth reading (theirs or not)	Link card (see the trade-off below)

Profile and Page posts are written differently: the profile is one person talking; the Page shows the team and the work, so people can reshare it with a sentence of their own. Never post the same thing to both on one day; give posts about the same subject a "topic" and schedule.py keeps them on different days.

Step 3: Write

• Hook in the first line, under about 140 characters (what a phone shows before "...see more"), with the specific thing in it: a number, a name, a result, a surprise. Line 2 earns the click on "see more" with a stake or a detail, not a date. Offer the user 2 or 3 first lines per post to pick from; the first line is the part worth their time.
• Then the substance: short paragraphs of whole sentences, first-hand detail, the user's words. Not one sentence per line ("broetry").
• End with a real question if the user has one, or just end. Never "comment YES", "like if you agree", "tag someone".
• No hashtag walls (0 to 3 that name the topic), Unicode bold or italic letters, stock phrases ("in today's fast-paced world", "game-changer", "let that sink in"), @names (the API can't tag people; write the name and tag in the app if it matters).
• Up to 3,000 characters; most good posts are far shorter.
• Links: a link card ("link") shows a preview of the page. Many practitioners measure less reach for posts with links; LinkedIn has never confirmed it, and putting the link in the first comment no longer seems to avoid it. Use a link when the click is the point; otherwise leave it out. For every link post, read the page first so the check can see redirects and the card image: python3 <skill>/scripts/page.py <url> --out work/page-<post>.json, then "page": "page-<post>.json".
• Page posts always carry a photo of people or the work, and a "reshare" list of 2 or 3 starter lines the team can rewrite in their own words; schedule.md lists them.

Patterns and examples: references/writing.md.

Step 4: Make the media

A real photo of the user, the team or the work beats any designed graphic. Use the user's files as they are when they fit (media in the plan); otherwise:

python3 <skill>/scripts/render_images.py work/images.json --out images/   # work/images/<id>.jpg, contact-sheet.jpg, manifest.json

Layouts (the spec format is at the top of the script): plain (the photo cropped to 4:5 or 1:1 around the subject), photo (with a short headline), quote, fact (one big number), list. 4:5 (1080x1350) gets the most room in the feed. Look at work/images/contact-sheet.jpg yourself. Image sets: 2 to 20 images in order, all the same shape (the first one sets the layout), one point each. Add "manifests": ["images/manifest.json"] to the plan.

Video from the user's own clip:

python3 <skill>/scripts/video.py fit clip.mov --ratio 4:5 --x 0.5 --start 2 --end 45 --out work/video/post.mp4
python3 <skill>/scripts/video.py captions work/video/post.mp4 --srt work/talk.srt --out work/video/post-captioned.mp4
python3 <skill>/scripts/video.py frames work/video/post-captioned.mp4 --out work/video/

Videos start muted in the feed: burn in captions from the user's own transcript (the posting API doesn't send LinkedIn's caption file). Open on the strongest moment; under 90 seconds unless the content needs more. 4:5 or 1:1 shows bigger than 16:9.

Document "carousels" (PDF slides) are a popular organic format, but WoopSocial doesn't post documents. Make the slides as images: post them as a 2 to 20 image set, or export them as a PDF for the user to upload in the app. Never say a document was scheduled.

Step 5: Check

python3 <skill>/scripts/li_check.py work/plan.json          # report, writes check.json
python3 <skill>/scripts/li_check.py work/plan.json --fix    # converted copies in fixed/, plan updated

It checks the text (length, the hook before the fold, bait, stock phrasing, Unicode bold, broetry, hashtags, @names, pasted URLs, the user's voice), the media (types, sizes, shapes, 2 to 20 images, video length and shape, never media and a link card together), the link card (full address, shorteners, redirects, a preview image), and the accounts (same post on profile and Page the same day, one a day, the week's count, repeats of a past post's picture or opening line). Run it until it exits with 0, and then deal with every warning: fix it, or tell the user and let them accept it. Warnings are where the reach and voice rules live.

Step 6: Pick the dates

python3 <skill>/scripts/schedule.py work/plan.json --start 2026-11-16

Each account posts on its own days and time, one a day at most, up to per_week (default 3), counting the log; posts on one subject (the same topic, text or picture) land on different days, on any account. It writes schedule.md (with each post's first-hour reply window and the days to check its numbers) and one WoopSocial request per post in requests/. Then run li_check.py again: with dates set, it checks the daily and weekly pace too.

Step 7: Upload and validate

Details: references/woopsocial.md.

1. Upload every file (media_uploads_create_session, <skill>/scripts/upload_parts.py, media_uploads_complete_session), one at a time. Put the IDs in each post's mediaIds (in order) and the accounts' socialAccountId in the plan, then python3 <skill>/scripts/schedule.py work/plan.json --bodies-only.
2. posts_validate every request. Fix anything it rejects and validate again.

Not connected and doesn't want to connect? Skip to the preview, hand over the posts, files and schedule, and stop. Never say something was scheduled when it wasn't.

Step 8: Preview, the yes, schedule

python3 <skill>/scripts/preview.py work/plan.json --out preview/   # work/preview/

Look at work/preview/preview.png yourself first: each post on a phone, cut at "...see more", with its image, grid, video or link card. If the point isn't above the fold there, rewrite the first line. Then give the user work/preview/index.html (the phone view and the full text) and work/schedule.md. Ask for one clear yes for the batch. Any change means a new check, new requests and a new preview.

After the yes: posts_create per validated request, write each post ID into the post's postId, posts_get for the status, then python3 <skill>/scripts/log.py record work/plan.json. Tell the user what's scheduled (post, account, date and time zone) and the first-hour plan in schedule.md.

Step 9: Learn from the numbers

From day 2 (most reach comes in the first day or two), and again at day 7:

• Company Pages: social_account_posts_list for the Page, saved as work/metrics.json.
• Profiles: LinkedIn doesn't give personal post numbers to posting apps. Ask the user to copy each post's analytics (impressions, reactions, comments, reposts, saves, sends, profile views, followers) into work/numbers.csv.

python3 <skill>/scripts/log.py results --plan work/plan.json --metrics work/metrics.json --csv work/numbers.csv

It compares each post with the same account's own posts: impressions against the median, then conversation (comments, reposts, saves, sends) per 1,000 impressions. A best post is named after 6.

What happened	Next
Impressions high, conversation low	People saw it but had nothing to say: end on a sharper point or a real question
Impressions low, conversation high	Good post, weak opening: rewrite hooks like this one more specifically
Both low	Change the subject or format before the wording
One format or subject wins	Do more of it, with new material

Quality bar

Before showing anything: would the best writer in the user's field post this, under their name? Reject and fix it if:

• the first line could open anyone's post;
• a claim, number or result can't be backed up;
• it reads like the stock AI phrasing list, or not like the user's own posts;
• it asks for engagement instead of earning it;
• profile and Page say the same thing the same day;
• the media is a generic graphic where a real photo would do.

Edge cases

• Several Pages: each Page is its own account in the plan, with its own days.
• Employees resharing a Page post: suggest they reshare with one sentence of their own; never write their posts for them as if they wrote them.
• Someone else's content: link to it or describe it; never repost their images or video. Public-domain or licensed material (NASA's photos, stock you paid for) is fine with a credit in the post, without implying the source endorses the user.
• AI images: LinkedIn shows a Content Credentials label when the file carries C2PA data, which re-rendering removes; post the generator's original file and say so in the text.
• A scheduled post needs changing (not yet live): delete it (posts_delete) and create it again, after confirming. Live: fix typos in the app; don't swap links or media.
• Polls, documents, articles, newsletters, comments, DMs: not through this skill. Prepare the text or file and the user posts it in the app.

References

• references/linkedin-specs.md: limits, formats, what the feed rewards and cuts, analytics, with sources.
• references/writing.md: hooks, structure, voice, Page posts, examples.
• references/plan-json.md: the plan file every script reads.
• references/woopsocial.md: connect, upload, validate, schedule, results.