new-project

Create New Data Science Project

Scaffold a new data science project using the copier-data-science template.

Prerequisites

• copier must be installed: pip install copier or uv tool install copier
• Template: https://github.com/stellarshenson/copier-data-science

Steps

1. ASK the user:

   • Project name (e.g. my-analysis)
   • Description (one line)
   • Author (name and email)
   • Python version (default: 3.12)
   • Location (default: current directory)

2. Run copier:

   copier copy https://github.com/stellarshenson/copier-data-science <project-name>
   

3. After scaffolding, seed the project .claude/CLAUDE.md (create the .claude/ dir if absent; if the template already wrote a CLAUDE.md, append the sections below under a ## Project working rules heading rather than overwriting):

   # Project: <project-name>
   
   Data science project scaffolded from copier-data-science. Extends the workspace / global configuration with the conventions this project runs by.
   
   ## Core engineering rules (precedence over everything below)
   
   1. **Think before coding** - state assumptions; surface tradeoffs; ask when unclear; present interpretations rather than silently picking one
   2. **Simplicity first** - minimum code that solves the problem, nothing speculative; if 200 lines could be 50, rewrite
   3. **Surgical changes** - touch only what the task needs; match existing style; remove only the orphans your change creates
   4. **Goal-driven execution** - turn each task into a verifiable goal (write the test / define the metric, then satisfy it); loop until verified
   
   ## Datascience plugin skills to use
   
   - `datascience:notebook-standards` - notebook structure, GPU-by-UUID, grouped imports, config render, rich output, figures, progress bars, checkpointing long runs
   - `datascience:hypothesis` - the experiments log + SOTA doc; run the project as falsifiable hypotheses (below)
   - `datascience:papers` - download + digest every cited paper into `references/papers/`
   - `datascience:progressbars` - a rich / tqdm progress bar for every medium or long loop
   - `devils-advocate:adversarial-review` - hostile review: data-scientist (experiment rigor), architect (project architecture), popular-science (the writeup), ux-designer (notebook visuals)
   - `datascience:prompt-engineering`, `datascience:footnotes` - prompt techniques; notebook / markdown footnotes
   - `datascience` - naming, file-format, and project-structure conventions (auto-applies)
   
   ## Run the project as hypotheses
   
   - Maintain a canonical append-only experiments log (`docs/experiments/<project>-experiments.md`) and a SOTA design doc (`docs/<project>-sota.md`) via `datascience:hypothesis`
   - Each hypothesis is `E<batch>-H<n>` with a 2-3 part memory slug (`E12-H37 graph-degree-lever`); refer by numeric id, add the slug where space allows; a batch may carry a focus slug (`E12 graph-theory-levers`)
   - Record a self-contained, independently reproducible experiment setup per hypothesis - a reader re-runs it from the doc alone, not from the transcript or by reading the code
   - Pre-register every hypothesis (prediction + falsifier / acceptance bar + diagnostic kill-gate) against a defined naive baseline; show the before / after summary as a markdown pipe table
   - Execute a hypothesis or a whole batch as a spawned agent on the selected model (best executor by default; ask when cost or scale warrants); record the execution model
   - Cite papers through `datascience:papers` (PDF + digest in `references/papers/`)
   - Before concluding SOTA, run an ablative study of the strongest hypothesis or all survivors to settle each component's marginal worth
   - Keep the executive summary and the research-at-a-glance table current every round