datascience

Hostile independent review by spawning fresh reviewer subagents - invokes the devils-advocate:adversarial-review skill, picking the adversary a data science project needs (data-scientist, architect, popular-science, ux-designer); find, fix, re-confirm clean

Add or fix footnotes in a notebook or markdown file using Jupyter-compatible anchor pattern

Add or fix progress bars in a notebook or script - choose classic (tqdm) or modern (rich) style

Apply rich output styling standards to a notebook or script - fix colors, formatting, print patterns

Apply the psychological prompting challenge - stakes, incentive, competitive framing. Use ONLY when the user explicitly asks for the challenge framing ("challenge me", "apply the challenge", "use competitive framing"); a hard problem alone is not the trigger, because the framing is a style the user opts into

Data science project conventions and standards. Auto-triggered when working with data science projects, notebooks, datasets, ML models, PyTorch, Polars, sklearn, or any data analysis workflow. Applies naming conventions, file format standards, project structure rules, and code patterns.

Acquire public and private datasets into a task folder under the project's external data directory - each corpus lands as a gitignored `dataset-*.zip` archive plus a tracked `dataset-*.md` sidecar recording licence, size, provenance, restrictions and the mapping onto the task, generated from one spec and re-rendered from the observed download so its row counts cannot drift. Use when fetching, downloading, adding, vendoring or documenting a corpus, benchmark or training set, when auditing what data a project holds and under what licence, or when a dataset arrives under access restrictions. Triggers - "download this dataset", "fetch the corpus", "add a dataset", "get the benchmark", "vendor this data", "what licence is this data", "document the dataset", "dataset sidecar", "external data".

Fix a notebook to comply with all standards - structure, styling, progress bars, header narrative

Port an existing project to copier-data-science standards or update an existing copier project

Markdown footnotes for Jupyter notebooks and markdown files using anchor links and span elements. Auto-triggered when adding references, citations, notes, or footnotes in any markdown context. Works in JupyterLab, GitHub, and standard markdown renderers.

Structure and maintain hypothesis-driven research documentation - a canonical append-only experiments log (each hypothesis with a self-contained, independently reproducible experiment setup, prediction, result, verdict) and a SOTA design doc distilling the winners. Use when the user is writing up an experiment, recording a hypothesis and its result, comparing approaches to decide which wins, defining a naive baseline, running an experiment's hypotheses as agents, fanning out the next round of hypotheses, ablating survivors into a final design, drafting a research report with a problem overview and executive summary, or concluding a state-of-the-art / final-design doc - even without the word "hypothesis". Triggers - "document this experiment", "write up the hypothesis", "experiments doc", "sota doc", "research writeup", "which approach won", "record this round", "run this experiment", "fan out hypotheses", "propose the next round", "ablation study", "update the experiments log", "structure my results".

Create a new data science project from copier template

Create a new Jupyter notebook with proper structure, styling, and progress bars

Jupyter notebook structure standards - section order, GPU-by-UUID selection, grouped imports, the configuration render, the semantic colour system (rich output + matplotlib colormaps), Polars-over-pandas dataframes, equations, charting every load-bearing result inline, fine-grained progress bars, checkpoint-and-restore for long runs, and the post-write compliance checklist. Use when creating or modifying a Jupyter notebook (.ipynb or Jupytext .py), when verifying or reviewing a finished notebook against the standards, or when the user mentions notebook structure, the config cell, GPU selection, rich output, colours / colour palette / colormap / plot styling, polars or pandas dataframes, notebook equations, matplotlib plots, charting or visualising results / conclusions / statistics, saving or exporting figures, progress bars, a slow or long-running cell, checkpointing / resuming / restoring an interrupted run, or a notebook checklist / standards check.

Paper reference workflow - download every cited paper, article, report or book and write a structured digest (overview, main findings, key takeaways, tags) carrying a resolvable online provenance link into the project's references/papers/ library. Use whenever a source is cited in a design / experiment / hypothesis / research document, recommended by a research agent, or handed over by the user - even without the word "paper". Triggers - "paper", "digest", "arxiv", "cite this paper", "reference this paper", "summarize this article", "digest this book", any citation added to project docs.

Use this skill when implementing progress bars in Python scripts or notebooks. Covers tqdm (classic) and rich (modern) styles, library configuration, Jupyter compatibility, and completion fixes.

Apply research-backed prompt engineering techniques to improve LLM output quality. Offers multiple techniques with templates and references. Auto-triggered when crafting system prompts, agent instructions, or LLM prompts.

Review a notebook or script for compliance with data science standards

Update a prompt, system instruction, or agent definition by applying a research-backed prompt engineering technique