scanpy

Overview of scanpy skill

What scanpy is for

The scanpy skill is for single-cell RNA-seq analysis in Python: quality control, normalization, feature selection, PCA, UMAP/t-SNE, clustering, marker discovery, and trajectory-style exploratory analysis. It is best when you already have or can create an AnnData object and want a practical scanpy guide for Data Analysis rather than a deep-learning model workflow.

Who should use it

Use the scanpy skill if you are working with scRNA-seq data, converting raw counts into an analysis-ready object, or preparing figures and cluster annotations for interpretation or publication. It fits analysts who want a standard exploratory pipeline and know they need scanpy install and usage steps, not a general-purpose bioinformatics primer.

What makes it different

This repository centers the common Scanpy workflow instead of a broad theory overview: data loading, preprocessing, embedding, clustering, and visualization. It also makes the boundaries clearer—if you need deep generative models, look to scvi-tools; if your main issue is AnnData structure or file formats, anndata is the better starting point.

How to Use scanpy skill

Install and verify the skill

Install the scanpy skill in the usual directory workflow, then open the skill file before prompting:

npx skills add K-Dense-AI/claude-scientific-skills --skill scanpy

After install, read scientific-skills/scanpy/SKILL.md first, then follow any linked sections inside the file. There are no extra helper scripts or reference folders here, so the skill text itself is the main source of truth.

Give the model analysis-ready inputs

The strongest scanpy usage prompts specify the data shape, format, and goal. Include whether your input is .h5ad, 10X output, or a count matrix; how many samples or batches you have; whether you need QC thresholds, integration, clustering, or plotting; and what output you want. A weak prompt says “analyze my scRNA-seq data”; a better one says “use scanpy to load two 10X datasets, filter low-quality cells, normalize, compute PCA/UMAP, cluster at multiple resolutions, and identify marker genes for each cluster.”

Follow a practical workflow

For most tasks, work in this order: load data, inspect obs/var, filter low-quality cells and genes, normalize/log-transform, select highly variable genes, scale and run PCA, build neighbors, compute UMAP, cluster, and then rank markers or visualize gene sets. If you already have processed AnnData, say so; that prevents the skill from repeating setup steps and keeps the output focused on the missing analysis decisions.

Read the right parts first

Start with the top of SKILL.md, then jump to the usage sections covering setup, loading, preprocessing, plotting, and downstream analysis. If the repository mentions specific input formats or conventions, match those exactly in your prompt. The goal is to reduce guesswork before the model writes code or analysis steps.

scanpy skill FAQ

Is scanpy only for scRNA-seq?

Yes, primarily. The scanpy skill is centered on single-cell transcriptomics workflows, especially exploratory analysis of gene expression matrices. It is not the right fit for bulk RNA-seq or for modeling tasks that are better handled by specialized single-cell deep-learning tools.

Do I need the skill if I can already prompt a model?

If you already know the full workflow and exact objects, a plain prompt may be enough. The scanpy skill is more useful when you want a consistent analysis path, fewer missing steps, and better defaults around QC, preprocessing, clustering, and visualization.

Is it beginner-friendly?

It is beginner-friendly in the sense that the pipeline is standard and the skill focuses on practical scanpy usage. It still assumes you understand basic Python, single-cell concepts, and the difference between raw counts, normalized data, and annotated metadata.

When should I not use it?

Do not use scanpy as the main skill if your problem is primarily file conversion, AnnData schema design, or building neural models for cell-state inference. In those cases, another skill or tool will give better results than forcing scanpy to cover every part of the workflow.

How to Improve scanpy skill

Provide the inputs that change decisions

The biggest quality gains come from stating dataset size, organism, platform, batch structure, and the analysis stage you are already in. For scanpy usage, also include whether you have ambient RNA concerns, doublet filtering needs, or a known set of marker genes, because those details affect QC and interpretation.

Ask for outputs, not just steps

Instead of asking for “scanpy analysis,” ask for a concrete deliverable: a preprocessing notebook, a clustering workflow, a marker-gene table, or a plotting block with UMAP and violin plots. Clear output targets help the model choose the right scanpy guide structure and avoid overexplaining basics you do not need.

Watch for the common failure modes

The most common mistakes are skipping QC context, mixing raw and normalized layers, clustering before checking neighbors and scaling choices, and asking for results without defining the biological question. If your first pass is too generic, revise with the exact input object, variable names, and the decision you want the analysis to support.

Iterate with one constraint at a time

After the first draft, improve scanpy skill output by adding one constraint: “use h5ad only,” “avoid integration,” “make publication-quality plots,” or “preserve batch labels in obs.” This is usually better than a broad rewrite because it keeps the workflow stable while tightening the parts that matter for your scanpy for Data Analysis task.