corvus �crow

A cookiecutter template for data-science, statistics, and text-analysis projects. Born as the internal template we use at Crow Intelligence for day-to-day analytics, modelling, and NLP research — and now available for anyone to use.

---

What you get

A project ready to work in, with no manual setup:

• Python at the version you choose, installed and pinned by uv (downloads a prebuilt python-build-standalone binary — no system build deps required)
• uv as package manager with a locked uv.lock
• ruff for linting and formatting + ty for type checking, both wired into pre-commit
• DVC for data versioning with a GCS remote
• MLflow with local tracking and GCS artifact store (optional)
• structlog for structured logging
• pydantic-settings + python-dotenv for typed config and GCS credential management
• Sphinx docs scaffold (autodoc, napoleon, RTD theme)
• Claude Code integration — a per-project CLAUDE.md and pre-installed skill packs (see below)
• Sensible .gitignore and .dvcignore defaults
• A Makefile with make help, make lint, make test, make dvc-push, make install-skills, and more

Project structure

my-project/
├── data/
│   ├── raw/          ← original, immutable data (DVC-tracked)
│   ├── processed/    ← transformed data (DVC-tracked)
│   └── external/
├── docs/             ← Sphinx docs
├── models/           ← serialised models and embeddings (DVC-tracked)
├── notebooks/        ← exploratory analysis
├── reports/figures/  ← generated graphics
└── src/my_project/   ← importable Python package
    ├── config.py     ← typed settings via pydantic-settings
    ├── logging.py    ← structlog setup
    └── tracking.py   ← MLflow helpers (if enabled)

---

Quickstart

uvx cookiecutter https://github.com/crow-intelligence/corvus.git

Or, if you don't have uv installed:

pip install cookiecutter
cookiecutter https://github.com/crow-intelligence/corvus.git

Follow the prompts. When done, cd into your new project and:

cp .env.template .env   # fill in your GCP credentials
uv sync
uv run dvc pull         # once data exists on the remote
make help               # see all available commands

---

Prompts

Prompt	Default	Notes
project_name	my-project	Slug and package name are derived automatically
python_version	3.11	Installed via uv python install — 3.11 resolves to the latest patch
licence	MIT	MIT, BSD 2/3, GPL/LGPL/AGPL v3, CC BY/BY-SA/BY-NC, Proprietary, Custom
gcs_bucket	gs://my-bucket	Used for DVC remote and MLflow artifacts
gcp_project_id	my-gcp-project	Required for GCS bucket operations
gcs_region	EU	Region for bucket creation
use_mlflow	yes	Adds MLflow and generates tracking.py
mlflow_experiment	project name	MLflow experiment name
use_spacy	no	Adds spaCy to runtime deps
install_claude_skills_python	yes	Vendor Matthew Honnibal's Python code-quality skills (MIT)
install_claude_skills_analytics	yes	Fetch nimrodfisher's 30 data-analytics skills at generation time
install_claude_skills_anthropic	no	Fetch Anthropic's official skill library (Apache-2.0)

---

Prerequisites

Before running corvus, you need:

• uv — package manager and Python installer (curl -LsSf https://astral.sh/uv/install.sh | sh)
• Google Cloud SDK — for DVC and MLflow on GCS (optional; you can configure GCS later)

---

Using MLflow

Tracking store is local (.mlruns/, gitignored). Artifacts go to GCS.

from my_project.tracking import init_experiment, start_run
import mlflow

init_experiment()

with start_run("feature-extraction"):
    mlflow.log_param("window_size", 5)
    mlflow.log_metric("coverage", 0.94)
    mlflow.log_artifact("reports/figures/vocab_growth.png")

View runs locally:

uv run mlflow ui

---

Using DVC

# After adding new raw data:
uv run dvc add data/raw/corpus.jsonl
git add data/raw/corpus.jsonl.dvc data/raw/.gitignore
git commit -m "data: add raw corpus"
uv run dvc push

# On another machine:
uv run dvc pull

---

Claude Code integration

Every generated project ships with:

• CLAUDE.md at the project root — per-project context (package name, Python version, layout, commands, MLflow/spaCy flags) that Claude Code loads automatically at session start.
• .claude/skills/ — pre-installed skill packs that Claude Code discovers on open.

Skill packs

Pack	Source	Default	Licence
python-quality	honnibal/claude-skills — vendored into this template	yes	MIT
data-analytics	nimrodfisher/data-analytics-skills — fetched at generation	yes	unspecified upstream
anthropic	anthropics/skills — fetched at generation	no	Apache-2.0

Flip any of these off at cookiecutter time. To change your mind later, edit .claude/skills/MANIFEST.yaml in the generated project and run make install-skills.

Scope

These skills are project-scoped — they live in the project's git repo and apply only when Claude Code is open in that project. For machine-wide skills, either symlink entries into ~/.claude/skills/ or install them there directly; project-level skills take precedence when both exist.

---

About Crow Intelligence

Crow Intelligence is an independent research group and boutique consultancy specialising in language, cognition, and AI. We apply computational methods to understand how language shapes thought — across historical corpora, political discourse, financial narratives, and beyond.

Corvus is the template we use for every new project. We're sharing it because good scaffolding shouldn't be reinvented each time.

---

Contributing

PRs are welcome — see CONTRIBUTING.md.