li-carousel

What does li-carousel do?

Humanize gate (required, every time)

No text from this skill is shown to the user until it has passed through the humanizer. No exceptions: short comments, one-line replies, headlines, hook options, "quick" drafts and revisions all go through it.

1. Write everything you're about to show into a temp file. Put a line with only --- between separate pieces (each option, comment, reply, slide, message).
2. Run the gate from this skill's folder (use python3 if python isn't found): python ../li-human/gate.py draft.txt -o clean.txt
3. If any piece is flagged or not PASS, rewrite those lines by hand in the user's voice and run the gate again. At most three rounds; if a piece still fails, show it but say which check failed.
4. Show the user only the text from clean.txt, never the pre-gate draft. Re-run the gate after every edit the user asks for.
5. End the reply with the gate's receipt line, exactly as printed, e.g. humanized: 3 pieces · 7 fixes · 86 PASS / SHORT PASS / SHORT PASS. A draft without a receipt line has not been through the gate.

If Python can't run here, say so plainly, apply ../li-human/slop.json by hand, and label the text not machine-checked. Never skip the gate silently.

Is it a carousel at all?

Carousels work when the idea has order or parts: steps, a framework, a before-and-after, a checklist, a ranked list. A single claim stretched over eight slides is a weak text post. If that's what the user has, say so and send them to /li-post.

Structure

8-12 slides. Fewer and it should be a text post; many more and people stop swiping.

1          COVER     the hook in 6 words or fewer, plus one line of promise
2          WHY       why this matters, one sentence
3..N       ONE IDEA PER SLIDE: 3-7 word headline, at most 25 words below it
N+1        SUMMARY   everything on one slide, so a screenshot is useful alone
LAST       CTA       one action: follow, comment, save, or visit a link

Copy rules

• The cover decides everything. Short, big, specific.
• Number the slides ("4/10"). Seeing the end keeps people swiping.
• No paragraphs. Over 25 words on a slide means two slides.
• The summary slide is the one people save. It must make sense alone.
• Handle on every slide, small, in a corner. Screenshots travel.
• Never invent numbers or results. Use {{number}} and ask.

Read ~/.claude/linkedin/voice.md if it exists; slide copy should still sound like the user.

Workflow

1. Write the slide copy as a numbered list the user can scan in seconds.
2. Write the caption: 2-4 lines that appear above the document in the feed. It needs its own hook (pick one from ../li-post/hooks.json).
3. Run the caption and slide copy through /li-human.
4. Wait for the user to approve the copy. Then build the PDF.

Building the PDF

LinkedIn accepts PDF documents. Use 1080x1350 px (4:5) pages; that fills the most screen on mobile.

Write one HTML file with a <section> per slide:

@page { size: 1080px 1350px; margin: 0; }
section { width: 1080px; height: 1350px; page-break-after: always;
          box-sizing: border-box; padding: 96px; }

• One accent colour, high contrast, nothing smaller than about 28px: people read these on phones.
• If the project has a brand guide or design system, use it rather than inventing one.
• Ask before using any logo or image the user didn't provide.

Print it to PDF with headless Chrome or Edge, for example:

chrome --headless --disable-gpu --no-pdf-header-footer --print-to-pdf=carousel.pdf carousel.html

(On Windows the binary is usually msedge or the full path to chrome.exe.) Check the page count matches the slide count, and open the PDF to confirm nothing overflows a page.

Output

Slide copy, caption, then the PDF path once built. Nothing is uploaded; the user posts the document themselves.