nicsuzor/extractor

Archive Information Extraction Skill

Assess archival documents (emails, correspondence, receipts) to identify information worth preserving in the knowledge base. This skill provides judgment criteria only - actual storage is handled by Skill(skill="remember").

Purpose

Filter noise from archival documents and identify genuinely significant information that should be preserved. Most archival documents have NO long-term value - be highly selective.

Assessment Criteria

Decision Framework

For each document, use LLM judgment to answer:

1. Is this important enough to mention at a monthly research meeting?
2. Would I want to remember this in 5 years?
3. Is this a concrete outcome, or just noise?
4. Did Nic take significant action, or is this passive/routine?

Trust your judgment - you understand context better than regex patterns.

What to Extract

1. Projects & Publications

• Papers: Actual submissions, acceptances, revisions, publications (not CFPs)
• Grants: Applications submitted, outcomes received, progress reports
• Books/Chapters: Contracts signed, submissions, publications
• Research Projects: Project initiation, significant milestones, new collaborations
• Awards: Significant awards or recognitions
• Milestones: Major career milestones, important advocacy/policy work
• Media: Media appearances or public engagement

Think: Is this a concrete action or outcome? Or just noise?

2. Professional Activities

• Events Organized: Conferences, workshops, panels Nic ran/chaired/coordinated
• Speaking: Confirmed talks, keynotes where Nic actually spoke
• Service: OSB decisions, editorial board work, advisory roles
• Reviews: Actual reviews submitted (not review requests)

Think: Did Nic DO something significant? Or just receive an invitation?

3. Applications & Career

• Grant Applications: Submitted applications (not drafts or ideas)
• Job Applications: Roles applied for
• Awards/Honors: Nominations received, awards won
• Promotion/Tenure: Applications, outcomes

Think: Is this a concrete milestone or application? Or just planning?

4. Important Contacts & Relationships

• Collaborations: Research collaboration proposals or agreements
• Partnerships: Grant writing partnerships, joint projects
• Networks: Connections made at conferences, international research networks
• Students: PhD supervision milestones, significant correspondence
• Institutions: New institutional relationships, key ongoing relationships

Think: Is this the start/milestone of an important relationship? Or routine correspondence?

5. Financial Records

• Receipts, invoices, contracts by project and category
• Log with: message ID, date, entity, project, description, amount, currency, payment status
• Essential for grant acquittal and tax purposes

Think: Will we need this for financial reporting?

What to Skip

Most documents are noise. Skip:

• Mass communications (newsletters, announcements, marketing, digests)
• Administrative routine (meeting scheduling, calendar invites, reminders)
• Declined invitations (conferences not attended, webinars skipped)
• Generic outreach (CFPs, generic collaboration requests, surveys)
• Automated systems (notifications, confirmations - unless significant)
• Spam/phishing
• Personal chitchat without professional substance
• Room bookings and logistics
• Normal student supervision (unless exceptional)
• Regular staff interactions
• Forum posts and mailing list discussions
• Course coordination and teaching logistics
• Standard reference requests

Test: Would I forget this within a week with no consequence?

Uncertain Cases

When uncertain:

• Lean toward extraction in initial pass
• Tag output with #review-classification for manual review
• Better to extract and discard later than miss important information

Judgment Examples

EXTRACT: "Your paper 'Platform Governance' has been accepted by Nature" → Clear publication outcome

SKIP: "CFP: Submit to Journal of Platform Studies by Dec 1" → Generic invitation, no submission

EXTRACT: "Congratulations, your FT210100263 grant has been awarded $500K" → Grant outcome with specifics

SKIP: "Reminder: Your FT210100263 annual report is due next month" → Administrative reminder

EXTRACT: "Following our chat at IGF, I'd love to collaborate on disinformation research..." → New substantive collaboration starting

SKIP: "You're invited to join our webinar on content moderation" → Mass invitation, not personal

EXTRACT: Email from examiner with detailed feedback on PhD thesis → Supervision milestone

SKIP: "HDR student seminars happening this week" → Generic announcement

EXTRACT: "OSB Case 2025R final decision: Upheld with modifications..." → Actual OSB work product

SKIP: "OSB weekly meeting reminder" → Routine scheduling

Key Information to Identify

When you decide to extract, identify these elements:

People

• Full name and affiliations
• Contact information (if provided)
• Role/position
• Relationship to Nic (collaborator, contact, student, etc.)

Organizations

• Universities and research institutions
• Research groups or labs
• Funding bodies
• Conferences or professional organizations

Events

• Conference/workshop name
• Speaking engagements
• Meetings with significance
• Research visits

Projects/Collaborations

• Research topics discussed
• Grant proposals
• Joint publications
• Ongoing or proposed collaborations

Dates & Timing

• When correspondence occurred
• When events happened or will happen
• Project timelines if mentioned

Topics/Tags

• Research areas (platform-governance, content-moderation, AI-ethics, etc.)
• Geographic locations
• Types of activity (collaboration, speaking, reviewing)

Identifiers (CRITICAL)

Always extract canonical identifiers to allow future reference:

• Email message ID (preferred)
• Source UUID or unique identifier
• URL or DOI
• If no canonical ID exists, create citation: "email from [sender] to [recipient] dated [YYYY-MM-DD] subject: [subject]"

Store identifier inline with extracted information.

Storage

Use Skill(skill="remember") to store extracted information in the knowledge base. The remember skill handles:

• Entity creation/updates
• Format validation
• Deduplication checking
• Proper categorization

Store information using the memory server with properly formatted markdown. Tag extracted information appropriately for searchability and future retrieval.

Common approaches for email extraction:

• Email metadata (sender, date, subject) → Use memory tags like #email, #correspondence
• Project information → Use memory tags like #project
• Coordinator/team roles → Store with descriptive relationships
• Timeline/duration → Use memory tags like #timeline, #deadline
• Status information → Use memory tags like #status
• Relationships between people → Document relationships clearly with role descriptions

Storage principles:

• Use clear, structured memory entries
• Include all relevant metadata and timestamps
• Tag appropriately for semantic search
• Avoid inventing custom categories or types
• Delegate all storage to the remember skill

Quality Standards

Confidence Levels

• Only extract when information is clear and valuable
• Mark uncertain extractions with #review-classification
• Skip ambiguous or unclear information

Atomic Information

• Each piece of information should be specific and verifiable
• Include dates, sources, and context
• Tag appropriately for searchability

Fail Fast

• If document is malformed, skip it and log error
• If extraction is ambiguous, skip it or tag for review
• Continue processing other documents