Social media kit

Every social post carries a video, GIF, or image. Native media is the largest reach lever the feed algorithms give you. This skill defines the kit of assets, the visual rules that keep a campaign from looking templated, and the browser mechanics for attaching media on each platform.

Configuration

Set these once. Scripts read them from environment variables, so nothing personal is hardcoded.

Setting	Env var	Default
Signature shown bottom-right on cards	BRAND_URL	yourblog.example
Output folder for cards	CARD_OUT_DIR	.
Screenshot for the runner card	CARD_SCREENSHOT	unset (card skipped)
Quote card attribution	QUOTE_ATTRIBUTION	Your Name · your role
Pass / fail dot colors	CARD_ACCENT_PASS, CARD_ACCENT_FAIL	3DD68C, F2555A
Banner accent color	BANNER_ACCENT	e13f5e (coral, configurable)
Fonts	HEADLINE_FONT, MONO_FONT	macOS Helvetica Neue and Menlo, DejaVu on Linux

Your handles, avatar and brand colors belong in your own config or prompt, never in this skill.

Requirements

• Python 3 with Pillow (pip install pillow)
• ffmpeg, for teaser cuts and clips
• A browser automation tool that can drive your logged-in Chrome over CDP (for example the browser-harness skill, if you have it)

The kit contract

Before posting, produce a kit folder (<scratchpad>/social-kit-<slug>/) with whichever of these the source material supports, in priority order:

Asset	Spec	Best for
teaser-16x9.mp4	30-60s cut of a master video, hook in the first 3s, H.264/AAC, captions burned in (feeds autoplay muted)	X thread root, LinkedIn
teaser-1x1.mp4	Same cut, 1080x1080 center-crop or pad	Mobile feeds
hero-card.png	1600x900 card with a padded screenshot or lane art	X, LinkedIn, Reddit image
terminal.gif	Animated terminal replay, 15MB or less (X GIF cap), 2-6 frames or a reveal sequence	X, Reddit
quote-card.png	1600x900 dark card with one strong line	Follow-up posts, LinkedIn comment

Reuse before creating. If your blog pipeline already produces per-post assets (OG gif, screenshots, a deep-dive mp4), a 30-60s ffmpeg cut of an existing master video beats a new render.

Style variety across a campaign

One visual language repeated on every asset reads as AI-generated. Rules:

• Never render every asset in a campaign in the same style. Rotate: lane-art card, terminal replica (see the terminal-replica skill), ASCII or OG gif, bare quote card, raw screenshot or recording.
• Match style to audience. Developer forums get terminal replicas or raw screenshots, because authentic beats polished. LinkedIn and X heroes get designed cards.
• PIL gotcha for terminal renders: Menlo lacks the TUI glyphs for the record dot, the elbow and the spinner. Draw them as vector shapes (filled circle, elbow lines), never ship tofu boxes.

House card style

Cards follow a dark thin-line art language, inspired by linear.app/now. The renderer is scripts/linearcard.py. The accent colors and signature are configurable.

• Canvas 1600x900 (render at 2x, LANCZOS downscale). Background #0A0B0D, flat. No gradients, no glow, no dot grids.
• Art is thin-line lane fields: horizontal 1.5px #43464C strokes flowing left to right, dots #9A9DA2 at start points, arrowheads at ends, lines breaking around thin circle nodes. Dashed lanes mean failures. The only accents are tiny green (pass) and red (fail) dots.
• Type: a bold sans for the big headline (white, 2-3 short lines, top-left). Mono for the eyebrow and data chips. Never set the headline in mono, and never bury the message in a caption.
• Real numbers in a chip row ("7 passed · 4 failed"), not decoration.
• Signature bottom-right: your URL in bold mono plus a cursor block. Nothing else.
• Screenshots sit on the art field with a 1px #303338 border and 20px rounded corners. Never full-bleed. Keep a headline beside them.
• Anything terminal-looking follows the terminal-replica skill.

BRAND_URL="myblog.dev" CARD_OUT_DIR=./kit python3 .claude/skills/social-media-kit/scripts/linearcard.py

The three cards in the script are examples. Edit the copy and numbers for your post.

Content type to visual treatment

Pick the treatment by what the post is, not by one default recipe.

Post is about	Reach for	Why
A flagship or launch idea	linearcard.py art field, but invent one concrete metaphor object for this post first (a shape that puns the idea)	A bespoke metaphor is more scroll-stopping than the same background every time
"Watch it actually run" proof (a real terminal session, an agent run)	cardrender.py screenshot or clip without the padded card frame, raw footage	A card frame reads as marketing polish. Raw footage reads as "this is happening"
A tool or stack story with numbers	linearcard.py standard recipe with a mono data-chip row	The data chips carry the story
A long essay excerpt or killer quote	quote-card.png, and drop the lane art entirely: pure type on the dark field	Background linework competes with a typographic excerpt

Thread media placement

In any multi-message thread (X thread, LinkedIn post plus comments, Reddit post plus comments), media goes on the first message only, preferably a gif or video. Later messages stay text-only unless an illustration adds information (a chart under discussion, a before/after). Decorating every reply with a card reads as botting and dilutes the root asset.

Production recipes

Teaser cut from a master video

# pick the strongest 30-60s arc: hook section plus one payoff section
ffmpeg -i master.mp4 -ss <hook_start> -to <hook_end> -c copy a.mp4
ffmpeg -i master.mp4 -ss <payoff_start> -to <payoff_end> -c copy b.mp4
printf "file 'a.mp4'\nfile 'b.mp4'\n" > cut.txt
ffmpeg -f concat -safe 0 -i cut.txt -c:v libx264 -crf 21 -pix_fmt yuv420p teaser-16x9.mp4
# 1:1 variant
ffmpeg -i teaser-16x9.mp4 -vf "crop=ih:ih,scale=1080:1080" -c:a copy teaser-1x1.mp4

If the teaser relies on narration, burn subtitles (-vf subtitles=teaser.srt:force_style='FontSize=22'). Muted autoplay kills uncaptioned voice-over.

Padded-card screenshots and cursor-animated clips

scripts/cardrender.py renders 1920x1080 frames: a screenshot floating on a gradient with rounded corners and shadow, an optional animated cursor with click ripples, row highlight pulses, and zoom. Use Seg(...) and render() for motion clips. For a static hero-card.png, render one frame:

from cardrender import Seg
seg = Seg("screenshot.png", 1.0, waypoints=[(0.5, 800, 400)], clicks=[], highlights=[], zoom=None)
seg.frame(0.5).save("hero-card.png")

Edit the gradient colors at the top of make_bg() to match your brand.

Terminal GIF

Author frames with PIL and a mono font on a dark background, reveal lines progressively, then img.quantize(colors=64) and save(..., save_all=True, duration=[...], loop=0). Keep under 15MB for X: 64 colors and 20 frames or fewer does it.

Terminal-style banner (article cover)

scripts/terminal_banner.py <out.png> draws a 5:2 (1500x600) terminal-window banner with a prompt line, a big wordmark with a block cursor, status lines and a feature list. Edit the CONFIG block per product. Use it when a raw app screenshot makes a poor cover.

Attaching media per platform

These mechanics assume a CDP-driven Chrome. Coordinates drift, so resolve elements through JS or getBoundingClientRect() each time. Screenshots can be a few percent larger than CSS pixels on retina displays, so never click straight from image coordinates.

Stream a file into the page

The raw DOM.setFileInputFiles call is the cleanest path when an input[type=file] exists (see below). When it does not work, stream the file in:

1. Base64 the asset, split it into 50KB chunks, and append each chunk with a js("window.__g += '<chunk>'") call. The JS pipe breaks above about 60KB.
2. In the page, build the file: atob, then Uint8Array, then new File, then DataTransfer.
3. Hand it to the composer as described per platform.

X

• Media, image or mp4: DOM.getDocument(depth=-1, pierce=True), then DOM.querySelector('input[data-testid="fileInput"]'), then DOM.setFileInputFiles(files=[path], nodeId=...). An 8MB 64s mp4 was ready in about 20s. Wait until the attachments strip stops showing upload progress.
• Fallback: a synthetic DragEvent('drop') (with dragenter and dragover first) on [data-testid="tweetTextarea_0"] works. Click the reply box first, wait about 2s, type the text, verify via innerText, then drop.
• Text entry: use CDP Input.insertText, not document.execCommand('insertText'). The draft editor silently mangles long multi-paragraph strings with execCommand. Focus the textarea via JS, call Input.insertText, then verify innerText (marker count, length range, URL present) before clicking Post.
• Self-reply chain: go to the parent status URL, focus tweetTextarea_0, insert text, verify, click tweetButtonInline by its rect, then find the new reply's status href by matching the first 40 characters of its text.
• Attach media to the thread root. Media plus a link in the same tweet suppresses the link card, which is fine: keep the canonical URL in the final reply, which then renders the card. Limits: mp4 2:20 or shorter, GIF 15MB, images 5MB. Screenshot-verify the media preview before clicking Post.
• Deleting your own tweet: the caret menu needs a real coordinate click, and it soft-rate-limits after several opens. Fallback is the web client's own delete GraphQL call from the page context with the session's csrf cookie. Look up the current query id in the web client's network log, because it changes.

LinkedIn

• Feed Photo and Video buttons both open a native file chooser. Use the file-chooser intercept below.
• Comments: drop does not work, and the visible input[type=file] elements belong to the messaging widget (attaching there sends it to a chat draft). Correct flow: patch HTMLInputElement.prototype.click to capture the input instead of opening the picker, click the comment box's "Add a photo" button (capture lands about 1s later), then set input.files = dt.files and dispatch change. Restore the prototype afterwards.
• Video post: set a thumbnail before Next (image icon under the preview, then "Add video thumbnail"). Otherwise the feed shows a black first frame.
• Native video beats a link preview card for reach. The URL can stay in the body text. 30-90s is the sweet spot.
• The create composer text box is reachable through .ql-editor, [contenteditable=true][role=textbox] with focus() plus Input.insertText. Some composers are closed shadow roots, so click by coordinates and verify by screenshot.
• LinkedIn rewrites links to lnkd.in shorteners. Verify an edit by curling the shortener, not by matching the raw URL.

Reddit

• The composer lives in shadow DOM: walk el.shadowRoot recursively. input[type=file][accept*=image] exists without clicking: assign files and dispatch input and change (bubbles and composed).
• The reopened composer restores old drafts. Clear it with real keys (click the editor, Select All, Backspace, verify innerText length 0) before typing, or new text interleaves with the stale draft.
• Read each subreddit's rules before posting. Image or video post types get more feed surface where allowed.

Native file-chooser intercept

Cleanest path when a real picker opens: cdp("Page.setInterceptFileChooserDialog", enabled=True), real-click the upload button, drain events, grab Page.fileChooserOpened's backendNodeId, call cdp("DOM.setFileInputFiles", files=[path], backendNodeId=bid), then disable the intercept. No base64 needed.

Hacker News and IndieHackers

HN is link-only: the OG image on the blog page does the work, so verify og:image renders before submitting. IndieHackers allows images in the body: include hero-card.png.

Facebook groups

Check the group's language and rules on its About page first, and confirm you are a member. Then:

1. Click "Write something..." and confirm the div[role=dialog] with a div[role=textbox] exists before typing. Closing the dialog preserves the draft.
2. Focus the textbox via JS and call Input.insertText.
3. The dialog already contains input[type=file]. Do not click Photo/video. Use DOM.querySelector('div[role=dialog] input[type=file]') plus DOM.setFileInputFiles with 2-4 images, and wait for blob <img> previews.
4. Click the div[aria-label="Post"] button. Posting shows a spinner for 15-25s in big groups.
5. Verify under the group's "my posted content" page. Large groups hold new posts as Pending (admin approval). Pending is success. Never repost while pending.

X Articles (long-form)

• Route: sidebar "Articles", then Write (not a direct compose URL). Drafts autosave server-side, so a browser restart loses nothing.
• Title: textarea[placeholder="Add a title"]. Input.insertText doubles text here, so set it with the native value setter plus an input event.
• Body: the single div[contenteditable=true]. Input.insertText works and preserves paragraphs.
• Cover (5:2, 1500x600): input[data-testid="fileInput"], then DOM.setFileInputFiles, then Apply in the crop dialog. Generate a banner with scripts/terminal_banner.py instead of using a raw app screenshot.
• Publish: Publish, then the audience dialog's Publish. The redirect to the article's status URL means it is live.

Session discipline

• Never delete an engaged post. Edit in place (LinkedIn: control menu, Edit post, which exposes a real .ql-editor).
• If you use the same Chrome the agent drives, the agent can re-attach to your focused tab between calls. Before any click batch, list tabs, switch to the target by URL, and assert location.href inside the same call. Do whole flows in one script, and prefer JS or API actions over coordinate clicks while you are working in the browser.
• Close leftover agent tabs before starting. Gate any Post click behind a content-verify loop that refuses to fire on a mismatch.

Post-shape rules that compound with media

1. Hook line above the media: one sentence, no link. The media stops the scroll, the first line earns the click.
2. One asset per post, root-only by default (see Thread media placement).
3. Match the asset to the claim. "Watch the agent work" gets the run clip. "One command" gets the terminal GIF.
4. Verify visually after posting. Screenshot each item and confirm the media rendered, because uploads fail silently during composer timeouts.
5. Reddit: verify with the JSON API. Moderator removal is invisible to the author: the post looks live while removed_by_category is moderator. Fetch /r/<sub>/comments/<id>.json and check removed_by_category and is_robot_indexable. A shadow-DOM file-input assign can also report success while attaching nothing, so check media_metadata on the live post JSON. If a subreddit removes your posts, move to a sibling community.
6. Keep your voice and content rules from your own publishing checklist: canonical URL everywhere, no marketing speak on Reddit, and hand off at auth walls.

Definition of done

• Kit folder produced, or an explicit note why no media exists for this topic
• Every posted item has native media attached and visually verified
• X thread: the root has the strongest asset and the final tweet carries the link
• Handoff files (HN, IndieHackers) reference the OG image or include hero-card.png