Run first-pass QC on a single-cell RNA-seq dataset

Drop a 10x Cell Ranger .h5 or an AnnData .h5ad file in front of Claude Code and get back a scverse-style MAD-filtered matrix plus the standard QC figures, ready to feed into Scanpy or scvi-tools.

   
Problem class Data analysis
Subject areas Molecular and Cellular Biology, Immunology and Microbiology, Neuroscience
Evidence level Reported
Complexity One skill or MCP
Availability Fully open
Compute Laptop

Problem

Every single-cell RNA-seq analysis starts the same way: filter empty droplets, drop dying cells with high mitochondrial fraction, remove obvious doublets, and produce the canonical violin / scatter QC figures so the next person who looks at the data can sanity-check the thresholds. The mechanics are well understood (the scverse “single-cell best practices” book is the reference), but writing the filtering boilerplate against AnnData / Scanpy each time is friction every group rebuilds — and the thresholds are easy to set wrong by hand. Solved looks like: hand the agent a file, get back a filtered AnnData object, a one-page QC report, and a record of the thresholds that were applied.

  1. Install the single-cell-rna-qc skill. From the anthropics/life-sciences marketplace:

    /plugin marketplace add anthropics/life-sciences
/plugin install single-cell-rna-qc@life-sciences

    Confirm with /plugin list. The skill ships with Python helpers that perform the MAD-based filtering described in scverse best practices.

  2. Put the input file next to your project. Either a 10x Cell Ranger filtered_feature_bc_matrix.h5 or an AnnData .h5ad. The skill auto-detects format.

  3. Invoke the skill in chat with the file path and any species-specific overrides. A minimal prompt:

    Run the single-cell-rna-qc skill on data/sample01.h5ad. Use the
default scverse MAD thresholds on n_genes, total_counts, and
pct_counts_mt. The species is mouse, so use mitochondrial gene
prefix "mt-" (not "MT-"). Write the filtered AnnData to
data/sample01_qc.h5ad and the QC figures to figures/sample01/.

  4. Review the QC figures and the threshold table the skill emits before passing the filtered AnnData to your downstream pipeline. If the thresholds look off (e.g., a tissue with genuinely high mitochondrial content like cardiomyocytes or hepatocytes), re-run with explicit mt_threshold and n_mads overrides — the SKILL.md documents the knobs.

  5. Hand off to Scanpy or scvi-tools. The filtered .h5ad is the entry point for Scanpy-MCP (interactive clustering / DE) or the scvi-tools skill (batch correction, deep-learning embeddings).

Why this assembly

Rung 2 of the simplicity ladder. Plain Claude Code can write Scanpy QC code from scratch each time, but every group then quietly drifts in how they set thresholds. The single-cell-rna-qc skill encodes the scverse MAD-based recipe so the thresholds are consistent and the figures are standardized; that is a meaningful gain over rung 1. No need to escalate to a toolbelt or an autonomous system — QC is a well-defined, single-step problem and a skill is the right grain.

Availability

Fully open. The skill is Apache-2.0-licensed and distributed via the anthropics/life-sciences plugin marketplace, free with any Claude plan. The underlying scverse stack (Scanpy, AnnData) is BSD-licensed. No subscription or institutional account required.

Compute requirements

Laptop-sufficient for typical datasets — a 10k-cell 10x sample QCs in a few minutes on a modern laptop with 16 GB RAM. For very large atlases (>500k cells), use backed mode and expect tens of minutes of wall-clock; the skill itself does not require a GPU.

Evidence

Reported. Anthropic’s tutorial for the single-cell-rna-qc skill walks through this exact assembly on an example 10x dataset, producing filtered output and QC plots. The skill was published as the first Anthropic-authored scientific skill alongside the Claude for Life Sciences launch in October 2025.

No peer-reviewed benchmark of “Claude + this skill” against a hand-coded Scanpy QC pipeline is known. The closest quantitative anchor is the scverse single-cell best-practices guide that the skill’s thresholds are drawn from; the agent loop adds convenience, not a new analytical method.

Alternatives considered

  • Plain Claude Code, no skill. Works — Claude can write the Scanpy QC code from scratch. Reach for this if you want to teach a student how the code works; reach for the skill if you want consistent thresholds across runs.
  • Scanpy-MCP alone. The Scanpy-MCP server exposes pp.* filtering primitives. Reach for it when QC is interleaved with clustering / DE in the same conversation. The dedicated skill is a better fit when QC is a one-shot pre-processing step run before downstream work.
  • An autonomous-science system (Biomni). Overkill for routine QC. Reach for Biomni only when QC is one node of a much larger autonomous pipeline (e.g., its published case study processing >336,000 single-nucleus profiles).

See also

Sources

Tried this recipe?

Share feedback — what worked, what didn’t, what you’d change. The form opens with this recipe pre-selected and a link back to this page.