starting-a-new-project

Starting a New Project

Rule: make the project understandable to a new RA and restartable by a future you.

Recommended layout

/gpfs/project/myproject/
├── code/                 # Git repo
│   ├── README.md
│   ├── CLAUDE.md
│   ├── pyproject.toml
│   ├── uv.lock
│   ├── renv.lock
│   ├── src/
│   ├── scripts/
│   └── slurm/
├── data/
│   ├── raw/              # read-only original data
│   └── derived/          # rebuildable intermediates
├── output/               # tables, figures, final outputs
├── logs/                 # Slurm logs
└── cache/                # API/download/model caches

First commands

For collaborative work, request a shared /gpfs/project/... folder from SOM IT instead of coordinating through one person's home directory. Put shared data, scripts, logs, and outputs there so permissions and ownership match the project.

mkdir -p /gpfs/project/myproject/{code,data/raw,data/derived,output,logs,cache}
cd /gpfs/project/myproject/code
git init
uv init --app
mkdir -p src scripts slurm

.gitignore

Track code, lockfiles, and documentation. Do not track data, credentials, environments, logs, or generated output. Pull a starting point from github/gitignore and add data/, logs/, output/, .env.

README minimum

# Project Name

## Setup

1. Clone repo into `/gpfs/project/myproject/code`.
2. Run `uv sync` on the login node.
3. If using R, run `Rscript -e 'renv::restore()'` on the login node.
4. Submit `sbatch slurm/test.sh`.

## Pipeline

1. `sbatch slurm/01_prepare_data.sh`
2. `sbatch slurm/02_estimate.sh`
3. Outputs appear in `/gpfs/project/myproject/output/`.

## Data

- `data/raw/`: original data, do not edit.
- `data/derived/`: rebuildable from raw data and code.

Seed a CLAUDE.md

Drop a short CLAUDE.md in the repo root so the project's cluster norms persist across sessions — Claude Code reads it automatically at the start of every session, and so does a future RA. Keep it to the project's specifics; the installed skills already carry the general rules.