# Conference Paper Roundup Skill

## When to trigger
Use this skill when the user asks to summarize papers from a conference,
workshop, or similar event — especially when papers are available as PDFs
on a programme page.

## Critical Pipeline: 4 Phases in Strict Order

NEVER start a later phase until the previous phase is 100% complete.

### PHASE 1: INVENTORY
1. Navigate to the programme page
2. Extract ALL paper titles, speakers, session names, and PDF URLs in one pass
3. Save to all_paper_links.json
4. Confirm count with user before proceeding

Rules:
- Scrape ALL days/sessions in one go
- NEVER fabricate URLs (conference sites use random hashes)
- Match papers to links by TITLE, not by position

### PHASE 2: DOWNLOAD
1. Use Python requests + pdfplumber to download and extract text
2. Save extracted text to extracted_texts/paper_NNN.txt
3. Report success/failure count

Rules:
- Use a single Python script, not multiple agents
- Extract first 5 pages only
- Handle .docx files with python-docx

### PHASE 3: SUMMARIZE
For each paper, create a JSON with:
- title, authors, authors_cite, authors_detail (name, institution, country)
- speaker (from PHASE 1 programme, NOT from the paper)
- abstract (verbatim), detailed_summary (2-3 paragraphs), one_liner
- topic, countries_of_study, is_rct, pdf_url

Rules:
- Speaker comes from the programme, not the paper
- Store 3 levels of detail to allow iteration
- Process in batches via Python script, not agents

### PHASE 4: VERIFY
1. Check for missing fields, wrong speakers, duplicates
2. Cross-check against programme ground truth
3. Produce verification report

## One-Liner Style
- Lead with the finding, not the method
- Include country/context
- Have personality
- End with (Author et al.) and #RCT if applicable

Good: "Phone surveys systematically distort responses by up to 63%
compared to in-person interviews in rural Nigeria (Castaing et al.). #RCT"

Bad: "This paper uses an RCT to study survey mode effects in Nigeria."

# Data Wrangling Skill

Default to R unless the user specifies otherwise.

## Core Principles
1. Print everything — every check produces visible output
2. Non-destructive first pass — flag before fixing
3. Report structure — numbered diagnostic report, pass/fail summary at top
4. Preserve provenance — log what changed, where, and why

## Diagnostic Checklist (in order)

### 1. Structure and Metadata
- Dimensions, variable names, types
- ID uniqueness (test explicitly)
- Duplicate rows and near-duplicates

### 2. Missing Values
- Count and percentage per variable
- Patterns: concentrated in certain groups/periods?
- Coded missing (-99, 999, "N/A") → convert and report
- Panel: differential missingness by treatment/round?

### 3. Outliers
- Summary stats with percentiles (1st, 5th, 95th, 99th)
- Statistical flags (3 SD, 1.5x IQR)
- Substantive flags (impossible values)
- Show flagged values in full-row context

### 4. Skip Patterns
- Identify filter variables and dependents
- Check values present when should be skipped
- Check missing when should be answered

### 5. Merge Diagnostics
- State expected cardinality (m:1, 1:1, m:m)
- Report match rates both directions
- Inspect unmatched observations
- Check for duplicated key inflation

### 6. Panel Consistency
- Balanced panel check
- Attrition analysis by key groups
- Time-invariant variables that vary

### 7. Treatment Variable
- Constant within unit across rounds
- Mutually exclusive and exhaustive
- Cross-tab with randomization strata