CLAUDE.md: Instructions file for Codex

AGENTS.md

This file provides global user preferences and guidance to AI coding agents when working with code on this system.

Software tools

Shell (mandatory)

All scripts should be runnable from the command line. The default shell of the user is fish, but scripts should be compatible with bash as well. All scripts are called from the root of the folder, and relative paths should be used to refer to files in subfolders.

Make (mandatory)

A single Makefile in the root of the folder should be used to run all scripts. The Makefile should have targets for each major step in the workflow, such as data wrangling, analysis, and reporting. Don't use Makefile as a script, but as a build tool: how to build the target from source using a recipe. For analytics, targets are data or research artifacts, sources are scripts and other data artifacts. For example, temp/clean_data.csv: code/create/clean.py input/raw/data.csv with the recipe python $<.

Bead (mandatory)

External data dependencies are managed by bead. Metadata is stored in .bead-meta/ subfolder. Input datasets are in input/ subfolder. Use bead input load <name> for missing data and bead input update <name> for updates. The temp/ subfolder is for temporary files, output/ for final output files. Check documentation at https://context7.com/bead-project/bead/llms.txt?tokens=10000

git (mandatory)

The folder is under version control with git. .gitignore should reflect the set of software tools used. Because bead is always used, .gitignore should always include input/ and temp/ folders. Avoid using git add . as this can include files that are not to be shared. Instead, add the files that have changed explicitly by name.

NEVER use git add -f to override .gitignore without explicit user permission. Files in .gitignore are there to:

• Protect sensitive data (credentials, private notes, proprietary data) • Exclude build artifacts and temporary files • Maintain repository hygiene