📋 CLAUDE.md: Teaching Claude Your Project

# CLAUDE.md — Electoral Fraud Detection Study

## Project Overview
Analysis of electoral fraud in African elections 2000–2020 using
digit-based statistical tests (Benford's Law) and machine learning.
Data from African Elections Database and NDI election observation reports.

## Technical Environment
- R version 4.3.2
- RStudio 2024
- Key packages: tidyverse, fixest, modelsummary, ggplot2, sf (spatial)
- renv is active — use renv::restore() to get all packages
- Do NOT use base R plotting (always ggplot2)
- Do NOT use dplyr::summarise() — use dplyr::summarize() (US spelling, per team convention)

## Folder Structure
R/            - All R scripts (numbered: 01_clean.R, 02_merge.R, etc.)
data/raw/     - Never touch
data/clean/   - Cleaned data (parquet format using arrow)
output/figs/  - ggplot figures (PDF + PNG)
output/tabs/  - modelsummary tables (LaTeX)
docs/         - RMarkdown reports

## Code Style
- Use tidyverse pipes (|>, not %>%)
- Variable names: snake_case
- Function names: snake_case with verb prefix (clean_votes(), merge_rounds())
- All scripts must start with: library(tidyverse); library(here)
- Use here::here() for all file paths — never setwd()

## Key Variables
- election_id: unique election identifier
- country_iso3: 3-letter country code
- fraud_score: our constructed fraud index (0–1 scale)
- turnout_pct: reported voter turnout (0–100)

## Rules
- Never use attach() or setwd()
- Always set seed (set.seed(42)) before any randomness
- Save figures at 300 DPI for publication quality
- Every output file must be reproducible from a fresh R session

# CLAUDE.md — Social Media Misinformation Study

## Project Overview
NLP analysis of COVID-19 misinformation on Twitter/X across
10 African countries. Dataset: 2.3M tweets, 2020–2022.
Goal: Classify misinformation, identify spreaders, measure policy impact.

## Technical Environment
- Python 3.11
- Virtual environment: venv (activate with: source venv/bin/activate)
- Key packages: pandas, numpy, scikit-learn, transformers, spacy
- GPU available: NVIDIA A100 (CUDA 12.1) — use when possible for model training
- Use pyproject.toml for dependency management

## Folder Structure
src/           - All Python source files
  data/        - Data loading and cleaning scripts
  models/      - ML model training scripts
  analysis/    - Analysis and evaluation scripts
data/raw/      - Raw Twitter data (never modify)
data/processed/ - Cleaned datasets (parquet format)
models/saved/  - Saved model checkpoints
notebooks/     - Exploratory Jupyter notebooks ONLY

## Code Style
- Follow PEP 8
- Type hints on all functions
- Docstrings on all functions (Google style)
- Use pathlib.Path, never os.path
- Use logging, never print() in production code

## Key Data Files
- tweets.parquet: main dataset (tweet_id, user_id, text, created_at, country)
- labels.csv: manual misinformation labels (tweet_id, label, confidence)
- users.parquet: user metadata

## Rules
- Never commit model weights to git (add to .gitignore)
- All experiments must be logged with MLflow
- Random seed: 42 everywhere
- Minimum test coverage: 80% for src/ directory