X Lookup

X lookup

One command gets you a JSON of posts (author, text, stats, image URLs, best-quality mp4 URLs) from any x.com link. No login needed for trends, single posts and threads. Don't rediscover this — run the script.

Fast path

node ~/.claude/skills/x-lookup/scripts/x-fetch.js "<x.com url>" --quiet --out <scratchpad>/x.json

Then read x.json (Read tool or node -e). On Windows, don't print it through Python without PYTHONIOENCODING=utf-8 — cp1252 chokes on emoji.

Flags:

Accepted URL kinds: x.com/i/trending/<id>, x.com/<user>/status/<id>, x.com/search?q=…, x.com/<user>. twitter.com and bare user/status/id are normalised.

Output shape

{
  "source": "...", "fetchedAt": "...",
  "title": "Claude Launches Projects for Managing Parallel AI Tasks",   // trend title or page title
  "updated": "Sep 17, 2026", "summary": "…Grok's trend summary…",       // trend pages only
  "loginWall": false,                                                   // true → rerun with --profile
  "count": 5,
  "posts": [{
    "id": "2100633571543367691", "url": "https://x.com/ClaudeDevs/status/…",
    "handle": "@ClaudeDevs", "name": "ClaudeDevs", "age": "1h",
    "text": "Today we're rolling out Projects…", "truncated": false,
    "images": ["https://pbs.twimg.com/media/XXXX?format=webp&name=orig"],
    "hasVideo": true, "poster": "https://pbs.twimg.com/amplify_video_thumb/…",
    "video": "https://video.twimg.com/amplify_video/…/3840x2160/….mp4",  // highest res seen
    "card": "https://external-link…", "quoted": "https://x.com/claudeai/status/…",
    "stats": { "replies": "205", "reposts": "192", "likes": "3.8K", "views": "251K", "duration": "00:25",
               "date": "10:46 PM · Sep 17, 2026", "bookmarks": "95" }   // date/bookmarks: focal post on a status page
  }]
}

Quirks to expect (don't treat as bugs):

• On a status page the quoted post and the replies come back as their own entries; the focal post is first.
• A quote-post that embeds a video shows the quoted post's video.
• poster for some videos is a generic pbs.twimg.com/media/… URL, not a thumb — the script already matches those by elimination.
• video URLs are sniffed from network requests while the page plays previews; if a video post has no video, run again with --full (the post page always loads the player).
• The sniffed video URL (…/vid/avc1/0/0/WxH/….mp4) is a DASH init segment, not a downloadable file — it's useful as an identifier and resolution hint only. To get the file use --download (yt-dlp) or hlsPlaylists[mediaId] with ffmpeg. Don't hand the raw URL to the user as "the video".
• Image URLs come back as format=webp&name=orig, which 404s if fetched directly — swap to format=jpg&name=orig (the download step does this).

Seeing what's in the media

When the post's substance is the video/image (a demo, a doodle reel, a screenshot), don't stop at the URL — look at it:

# images: just Read the downloaded .jpg
# video: contact sheet, one frame every N seconds, then Read it
ffmpeg -v error -y -i DIR/<file>.mp4 -vf "fps=1/6,scale=480:-1,tile=4x3" DIR/contact.jpg
ffprobe -v error -show_entries format=duration:stream=width,height -of csv=p=0 DIR/<file>.mp4

Pick the fps so 12 tiles cover the whole clip (1/6 for ~70s, 1/2 for ~25s). For a specific moment: -ss 12 -frames:v 1. Needs ffmpeg on PATH (see README). Describe the frames concretely — what's on screen, on-screen text, the visual gag — that's what the user wants when they ask "what's this post".

Reporting back

Lead with the trend title + one-line summary, then a compact table: who · text (trimmed) · media (🎬 duration / 🖼 count) · stats. Put direct media URLs below the table when the user is likely to want them (reels, receipts, "get the video"). Quote text verbatim — the user cares about exact wording.

Why this path

• WebFetch / plain HTTP on x.com → 402 or the empty React shell. X only renders through a real browser.
• Headless Chromium + a desktop UA is enough: X serves a full logged-out render for trends, posts and threads.
• The script finds Playwright (local, global npm, or the copy bundled in @playwright/mcp) and the newest Chromium in Playwright's browser cache on Windows, macOS and Linux — no config needed.
• Logged-out DOM has no data-testid — the script parses the text layout (name / @handle / age / body / [m:ss] / replies reposts likes views). Logged-in DOM has testids and the script prefers them.
• waitUntil: 'load' never fires on X; the script uses domcontentloaded + waitForSelector('article').

When it's not enough

• Need more than the top ~10 posts of a trend, or search, or a profile timeline → ask the user to run --login once, then add --profile.
• Need replies/thread depth → pass the status URL; scroll count controls how many replies come back.
• Media for a reel → --download into the reel's assets dir, then hand off to /media-use / the reel skill.