Extract Insights from Reference Collections

When This Activates

• "Extract insights from my Are.na"
• "Synthesize these articles"
• "Turn my bookmarks into something useful"
• "What patterns are in my saved references?"

The Process

Phase 1: Audit

Understand the collection contents.

AUDIT CHECKLIST:
- Total items: [count]
- By type: [links, images, text, attachments]
- Link domains: [group by source - medium.com, specific blogs, etc.]
- Blocked sources: [paywalled, bot-detected]
- High-value targets: [articles > images > text snippets]

For Are.na: Run npx tsx scripts/audit-channel.ts <slug> if available, or fetch via API.

Priority order for insight extraction:

1. Articles/links (richest content)
2. Text blocks (direct notes)
3. Images with descriptions (frameworks, diagrams)
4. Images without descriptions (least context)

Phase 2: Deep-Read

For articles/links:

FETCH STRATEGY:
1. Direct fetch (WebFetch)
2. If blocked or inaccessible:
   - Flag for manual review
   - Note: "[url] - blocked, needs manual access"
3. If success:
   - Extract key insights (not full content)
   - Note source for citation

For PDFs and attachments:

PDF STRATEGY:
1. Check file size first (if available from metadata)
   - < 5MB: Read directly with Read tool
   - 5-20MB: Read first 50 pages only
   - > 20MB: Skip, flag for manual review
2. If no size available, attempt read:
   - If succeeds: extract insights
   - If times out or fails: skip and flag
3. For scanned PDFs (no extractable text):
   - Skip, flag as "scanned PDF - needs OCR"

Skip criteria (don't waste time):

• PDFs > 20MB (likely books or massive reports)
• Scanned documents without OCR
• Zip files or archives
• Video/audio files (note title only)

For each article, extract:

SOURCE: [title] - [url]
TYPE: [framework / case-study / opinion / tutorial / research]
KEY INSIGHT: [1-2 sentence core idea]
ACTIONABLE: [Yes/No - can this become a process?]
FRAMEWORKS: [Any named methodologies?]

Phase 3: Cluster

Group insights by theme:

CLUSTERING QUESTIONS:
- What problems do multiple sources address?
- What methodologies appear repeatedly?
- What contradictions exist?
- What's unique vs. common knowledge?

Output format:

THEME: [name]
Sources: [list]
Core insight: [synthesis]
Skill potential: [High/Medium/Low]

Phase 4: Synthesize

For high-potential themes, create actionable output:

SYNTHESIS OPTIONS:
1. Skill - if it's a methodology Claude should auto-apply
2. Prompt - if it's a one-time process user would paste
3. Reference doc - if it's knowledge, not process
4. Discard - if it's common knowledge or too vague

Skill criteria:

• Specific trigger conditions
• Named principles or laws (not invented)
• Clear diagnostic/output format
• Adds value over Claude's base knowledge

Output Format

INSIGHT EXTRACTION REPORT

COLLECTION: [name/url]
AUDITED: [date]
TOTAL ITEMS: [count]
SUCCESSFULLY READ: [count]
BLOCKED: [count]

THEMES IDENTIFIED:
1. [Theme]: [X sources] - [skill potential]
2. [Theme]: [X sources] - [skill potential]

HIGH-VALUE INSIGHTS:
- [Insight] (from: [source])
- [Insight] (from: [source])

RECOMMENDED ACTIONS:
- Create skill: [name] - based on [theme]
- Create prompt: [name] - for [use case]
- Manual review needed: [blocked URLs]

BLOCKED SOURCES (need manual access):
- [url] - [reason]

SKIPPED (too large or unsupported):
- [filename] - [size] - [reason]

Failure Modes

When to Stop

Stop when sufficient insights are extracted:

• You've extracted 10+ actionable insights
• Remaining sources are blocked/inaccessible
• Themes are repeating (saturation)
• User has enough to work with

Extract what's accessible, flag what's not, move on.