Categorizing Bsky Accounts

Categorizing Bluesky Accounts

Fetch Bluesky account data and extract keywords for Claude to categorize by topic. The script compresses account context (bio + posts) into bio + keywords, then Claude performs intelligent categorization.

Prerequisites

Requires: extracting-keywords skill (provides YAKE venv + domain stopwords)

The analyzer delegates keyword extraction to the extracting-keywords skill, which provides:

• Optimized YAKE installation with minimal dependencies
• Domain-specific stopwords: English (574), AI/ML (1357), Life Sciences (1293)
• Support for 34 languages

Core Workflow

When users request Bluesky account analysis:

1. Ensure keyword extraction is set up - Invoke the extracting-keywords skill using the Skill tool to ensure YAKE venv exists (skip if already invoked in this session)

2. Determine input mode based on user's request:

   • Following list → use --following handle
   • Followers → use --followers handle
   • List of handles → use --handles "h1,h2,h3"
   • File provided → use --file accounts.txt

3. Configure parameters:

   • --accounts N - Number to analyze (default: 100, max: 100)
   • --posts N - Posts per account (default: 20, max: 100)
   • --stopwords [en|ai|ls] - Choose domain-specific stopwords:
     • en: English (general purpose)
     • ai: AI/ML domain (recommended for tech accounts)
     • ls: Life Sciences (for biomedical/research accounts)
   • --exclude "pattern1,pattern2" - Skip spam/bot accounts

4. Run script - Outputs simple text format to stdout:

   @handle1.bsky.social (Display Name)
   Bio text here
   Keywords: keyword1, keyword2, keyword3
   
   @handle2.bsky.social (Another Name)
   Bio text here
   Keywords: keyword4, keyword5, keyword6
   

5. Categorize accounts - Claude analyzes bio + keywords to categorize by topic

Quick Start

Analyze following list with AI/ML stopwords:

python scripts/bluesky_analyzer.py --following austegard.com --stopwords ai

Analyze followers:

python scripts/bluesky_analyzer.py --followers austegard.com

Analyze specific handles:

python scripts/bluesky_analyzer.py --handles "user1.bsky.social,user2.bsky.social,user3.bsky.social"

From file:

python scripts/bluesky_analyzer.py --file accounts.txt --stopwords ai

Filter out bot accounts:

python scripts/bluesky_analyzer.py --following handle --exclude "bot,spam,promo" --stopwords ai

Parameters

Input Modes (choose one)

--handles "h1,h2,h3" Comma-separated list of Bluesky handles

--following HANDLE Analyze accounts followed by HANDLE

--followers HANDLE Analyze accounts following HANDLE

--file PATH Read handles from file (one per line)

Analysis Options

--accounts N Number of accounts to analyze (1-100, default: 100)

--posts N Posts to fetch per account (1-100, default: 20)

--stopwords [en|ai|ls] Stopwords to use for keyword extraction (default: en)

• en: English stopwords (574 terms) - general purpose
• ai: AI/ML domain stopwords (1357 terms) - tech-focused accounts
• ls: Life Sciences stopwords (1293 terms) - biomedical/research accounts

--exclude "word1,word2" Skip accounts with these keywords in bio/posts

Output Format

The script outputs simple text format for Claude to process:

@alice.bsky.social (Alice Smith)
AI researcher working on LLM alignment and safety
Keywords: alignment, safety research, interpretability, llm evaluation

@bob.bsky.social (Bob Johnson)
Full-stack developer building web applications
Keywords: react, typescript, node.js, api design, postgresql

@carol.bsky.social (Carol Williams)
Biotech researcher studying CRISPR applications
Keywords: crispr, gene editing, therapeutics, clinical trials

Claude then categorizes accounts based on bio + keywords without hardcoded rules.

Common Workflows

Audit Your Following List

python scripts/bluesky_analyzer.py --following your-handle.bsky.social --stopwords ai

Claude will categorize accounts by topic and identify patterns in who you follow.

Find Experts in a Topic

python scripts/bluesky_analyzer.py --following alice.bsky.social --stopwords ai

Ask Claude: "Which of these accounts are ML researchers?" or "Who focuses on climate tech?"

Analyze a Curated List

cat > accounts.txt << 'EOF'
expert1.bsky.social
expert2.bsky.social
expert3.bsky.social
EOF

python scripts/bluesky_analyzer.py --file accounts.txt --stopwords ls

Filter Out Bot Accounts

python scripts/bluesky_analyzer.py --following handle --exclude "bot,spam,promo,follow back" --stopwords ai

Technical Details

Keyword Extraction

Delegates to extracting-keywords skill using YAKE venv:

• Stopwords options (--stopwords):
  • en: English (574 terms) - general purpose
  • ai: AI/ML domain (1357 terms) - filters technical noise, ML boilerplate
  • ls: Life Sciences (1293 terms) - filters research methodology, clinical terms
• N-grams: 1-3 words
• Deduplication: 0.9 threshold
• Top keywords: 10 per account
• Performance: ~5% overhead with domain stopwords vs English

API Rate Limits

Bluesky API limits:

• 3000 requests per 5 minutes
• 5000 requests per hour

The analyzer respects these limits with built-in delays.

Categorization Algorithm

Script's role:

1. Fetch account data (bio + posts)
2. Extract keywords to compress context
3. Output bio + keywords in simple format

Claude's role:

1. Read bio + keywords for each account
2. Intelligently categorize by topic (no hardcoded rules)
3. Group accounts, identify patterns, answer user questions

This agentic pattern is more flexible than hardcoded keyword matching.

Troubleshooting

"No accounts to analyze"

• Verify handle format (include domain: handle.bsky.social)
• Check if account exists and has public following/followers

"Insufficient content for keyword extraction"

• Account has few posts (<5)
• Posts are very short
• Try increasing --posts parameter

Rate limit errors

• Reduce --accounts parameter
• Add delays between batches
• Check Bluesky API status

Import errors

• Verify extracting-keywords skill is available
• Check YAKE venv exists: /home/claude/yake-venv/bin/python -c "import yake"
• Verify Python 3.8+: python3 --version

Integration with Other Skills

Built-in integration:

• extracting-keywords: Automatically delegates keyword extraction to this skill's optimized YAKE venv with domain-specific stopwords

Categorizing Output

Group accounts by topic from their bio + keywords, using judgment rather than a fixed taxonomy — the categories and counts follow from what's actually in the data, not a template.