x

name	x
description	Automate X (Twitter) for research, video understanding, and content creation. Use when navigating X, searching posts, analyzing videos, or creating content.
allowed-tools	Bash

X Automation

Quick Reference

Task	Trigger	Description
Navigate	"open x.com"	Open X, verify login
Search	"research posts about"	Search keyword, Top/Latest
Download video	"download video from"	Download with yt-dlp
Understand video	"understand this video"	Analyze with frames or audio
Compose	"create post about"	Draft content
Post	"post it"	Publish (requires approval)

Navigate to X

When asked to open X or navigate to Twitter:

1. Get browser tab: mcp__claude-in-chrome__tabs_context_mcp(createIfEmpty: true)
2. Create new tab: mcp__claude-in-chrome__tabs_create_mcp()
3. Navigate: mcp__claude-in-chrome__navigate(url: "https://x.com", tabId: <id>)
4. Wait for load: mcp__claude-in-chrome__computer(action: "wait", duration: 3, tabId: <id>)
5. Screenshot to verify: mcp__claude-in-chrome__computer(action: "screenshot", tabId: <id>)
6. Check for compose button = logged in

Error: If login page appears, stop and inform user to log in manually.

Search Posts

When asked to search or research posts:

1. Find search box: mcp__claude-in-chrome__find(query: "search bar", tabId: <id>)
2. Click it: mcp__claude-in-chrome__computer(action: "left_click", ref: <ref>, tabId: <id>)
3. Type keyword: mcp__claude-in-chrome__computer(action: "type", text: "$ARGUMENTS", tabId: <id>)
4. Submit: mcp__claude-in-chrome__computer(action: "key", text: "Enter", tabId: <id>)
5. Wait: mcp__claude-in-chrome__computer(action: "wait", duration: 2, tabId: <id>)
6. Switch to Latest: mcp__claude-in-chrome__find(query: "Latest tab") → click
7. Read results: mcp__claude-in-chrome__read_page(tabId: <id>)

Read Post Content

When navigating to a specific post URL:

1. Navigate to post URL
2. Wait for load: mcp__claude-in-chrome__computer(action: "wait", duration: 2, tabId: <id>)
3. Take screenshot: mcp__claude-in-chrome__computer(action: "screenshot", tabId: <id>)
4. Read DOM: mcp__claude-in-chrome__read_page(tabId: <id>)
5. Extract: author, text, timestamp, media type, engagement metrics

Download Video

When a post contains video and you need to understand it:

yt-dlp -o "downloads/%(id)s.%(ext)s" "<post_url>"

If download fails, try with browser cookies:

yt-dlp --cookies-from-browser chrome "<post_url>"

Check Video Audio

After downloading, check for audio:

ffprobe -v error -select_streams a -show_entries stream=codec_type -of csv=p=0 "<video_path>"

• Output contains "audio" → Has audio, use transcription
• Empty output → No audio, use frame extraction

Understand Silent Video (No Audio)

1. Get duration:

ffprobe -v error -show_entries format=duration -of default=noprint_wrappers=1:nokey=1 "<video>"

2. Extract frames every 3 seconds:

ffmpeg -i "<video>" -vf "fps=1/3" "frames/frame_%04d.png"

3. Read each frame visually with Claude

4. Synthesize what the video shows

5. Cleanup:

rm frames/*.png

Understand Video with Audio

1. Extract audio:

ffmpeg -i "<video>" -vn -acodec mp3 "audio/temp.mp3"

2. Transcribe with Whisper:

whisper "audio/temp.mp3" --model base --output_format txt --output_dir "transcripts/"

3. Read transcript and summarize content

4. Cleanup:

rm "audio/temp.mp3"

Compose Post

When creating content:

1. Find compose button: mcp__claude-in-chrome__find(query: "compose tweet button", tabId: <id>)
2. Click it: mcp__claude-in-chrome__computer(action: "left_click", ref: <ref>, tabId: <id>)
3. Wait: mcp__claude-in-chrome__computer(action: "wait", duration: 1, tabId: <id>)
4. Type content: mcp__claude-in-chrome__computer(action: "type", text: "<content>", tabId: <id>)

If uploading media: 5. Find media button: mcp__claude-in-chrome__find(query: "add photo", tabId: <id>) 6. Click it and upload file 7. Wait for upload: mcp__claude-in-chrome__computer(action: "wait", duration: 3, tabId: <id>)

CRITICAL: Do NOT click post button yet. Show user the draft first.

Post (Requires Approval)

NEVER post without explicit user approval.

1. Show user the composed content
2. Ask: "Should I post this to X?"
3. Wait for explicit "yes" or approval
4. Only then: find post button and click

Upload Image to Post

1. Ensure compose dialog is open
2. Find media button: mcp__claude-in-chrome__find(query: "media upload", tabId: <id>)
3. Click: mcp__claude-in-chrome__computer(action: "left_click", ref: <ref>, tabId: <id>)
4. Find file input in DOM: mcp__claude-in-chrome__read_page(filter: "interactive", tabId: <id>)
5. Set file: mcp__claude-in-chrome__form_input(ref: <file_input>, value: "<image_path>", tabId: <id>)
6. Wait for upload preview: mcp__claude-in-chrome__computer(action: "wait", duration: 3, tabId: <id>)
7. Verify image appears in compose box

Analyze Post Engagement

When researching what's performing well:

1. Read post metrics from DOM
2. Extract: replies, reposts, likes, views, bookmarks
3. Calculate engagement rate
4. Identify high-performing content patterns

Extract Trending Topics

When researching trends:

1. Search for keyword
2. Read multiple posts (5-10)
3. Identify common themes, hashtags, phrases
4. Summarize trending topics for content creation

Generate Content Ideas

Based on research:

1. Analyze trending topics from research
2. Identify humor patterns, meme formats in use
3. Generate 2-3 content ideas
4. Present to user for selection

Create Meme Concept

For meme creation:

1. Based on selected idea, describe meme concept:
   • Top text
   • Bottom text
   • Image description
2. Use external tool (Gemini/Nano Banana Pro) to generate image
3. Save to output directory
4. Present to user for approval before uploading

Composite Workflows

Research and Post Workflow

Trigger: "research X and post about it"

1. Navigate to X
2. Search for keyword (Latest tab)
3. For each post: read content, check for video, analyze engagement
4. Extract trending topics
5. Generate content ideas
6. Present ideas to user
7. Create meme concept
8. Compose with content
9. Upload meme image
10. GET USER APPROVAL
11. Post

Quick Research Workflow

Trigger: "what's trending about X"

1. Navigate to X
2. Search for keyword (Top tab)
3. Read top 5 posts
4. Summarize findings
5. Return summary (no posting)

Video Understanding Workflow

Trigger: "understand the video in this post"

1. Navigate to post URL
2. Download video with yt-dlp
3. Check if video has audio
4. If audio: extract and transcribe with Whisper
5. If no audio: extract frames every 3 seconds
6. Analyze content (transcript or frames)
7. Summarize key takeaways
8. Cleanup temporary files

Safety Rules

1. NEVER auto-post without explicit user approval
2. Always show composed content before posting
3. Log all posting actions
4. Respect X rate limits - pause if limited
5. Read-only by default - posting requires explicit command

Error Recovery

Page Not Loading

• Retry navigation 3 times
• Increase wait duration each attempt

Element Not Found

• Try alternative query terms
• Scroll page and retry
• Take screenshot to diagnose

Login Required

• Stop immediately
• Inform user: "Please log into X manually, then retry"

Video Download Failed

• Try with --cookies-from-browser chrome
• Check if post is accessible
• Inform user of failure

Whisper Failed

• Fall back to visual frame analysis
• Inform user transcription unavailable