K-Dense-AI/claude-scientific-skills – A set of ready to use Agent Skills for research, science, engineering, analysis, finance and writing.

K-Dense-AI/claude-scientific-skills: a skills library for scientific Claude use

I have been using claude-scientific-skills (a community-maintained collection of reusable skill definitions for Anthropic’s Claude, focused on scientific computing tasks: data analysis, plotting, numerical methods, literature search, and reproducibility tooling) for about three months. The repository is one of the more interesting things I have seen in the “AI for science” space, because it does not try to be a model. It is a set of capabilities you can hand to a model that already exists.

The premise is simple. Claude is a general-purpose assistant. Scientific computing has specific needs: a way to load a CSV, a way to fit a curve, a way to make a publication-quality plot, a way to cite a paper. The skills library is a curated set of “skills” (small bundles of code, prompts, and documentation that Claude can use to perform a specific task) that wrap the general-purpose model with the specific capabilities a scientist needs.

This is a writeup of what the library does, what it is good for, and where it falls short.

What a “skill” actually is

A skill in the library is a directory with a few files.

• A SKILL.md file. A markdown file that describes what the skill does and how to use it. The file is loaded into Claude’s context (the text the model is given to “remember” for the current conversation) when the skill is activated.
• A set of helper scripts. Python or R scripts that perform the common operations the skill supports: load a data file, fit a model, generate a plot.
• A requirements.txt or equivalent. The dependencies the helper scripts need.

The model is given the skill description and the helper scripts. When the user asks for a task that matches a skill, Claude uses the skill’s scripts to do the work. When the user asks for a task that does not match a skill, Claude falls back to its general capabilities.

The pattern is the same as the one used by Claude Code (Anthropic’s CLI for Claude, which lets you run Claude from a terminal with the same context and tool system as the web interface) and other agent systems. The library’s contribution is the scientific domain expertise baked into the skill definitions.

What is in the library

The library has about 30 skills as of early 2026. The categories are the ones that come up in scientific computing.

• Data loading and cleaning. Skills for common scientific data formats: CSV, TSV, NetCDF (a self-describing binary format widely used in climate science and geophysics), HDF5 (the Hierarchical Data Format version 5, a binary format used to store large numerical datasets), and a few instrument-specific formats. The skills handle the boilerplate of reading the data into a usable in-memory structure.
• Statistical analysis. Skills for descriptive statistics, hypothesis testing, regression, and the more common multivariate methods (PCA, which stands for Principal Component Analysis, a technique for reducing the dimensionality of a dataset; clustering, which groups similar data points together; and ANOVA, which stands for Analysis of Variance, a way to test whether group means differ significantly). The skills are not a replacement for a real statistics textbook. They are a way to do the common operations without writing the code from scratch.
• Plotting. Skills for publication-quality plots: matplotlib for Python, ggplot2 for R. The skills include the boilerplate for common plot types and a style template that produces plots that look reasonable in a paper.
• Numerical methods. Skills for root finding, integration, ODE (ordinary differential equation, an equation relating a function to its rate of change) solvers, and optimization. The skills wrap scipy and nlopt (an open source library for nonlinear optimization) with a consistent interface.
• Literature search. A skill for searching arXiv, PubMed, and a few other academic databases. The skill formats the queries, retrieves the results, and returns a structured summary.

The library is not exhaustive. The categories are the ones the maintainers have written skills for. The list grows when someone contributes a new skill. The growth rate is about one new skill per month.

What it is good for

I want to be specific about the use cases I have found the library useful for.

• Exploratory data analysis. The data loading, statistical analysis, and plotting skills are a real time saver for the first hour of looking at a new dataset. I can ask Claude to “load this CSV, give me summary statistics, and plot the distribution of column X” and get a useful answer in about 30 seconds.
• Teaching. I have used the library to help students learn scientific computing. The skill structure makes it easy to point at a specific skill and say “this is how you fit a linear model.” The skills are a teaching tool, not a replacement for understanding.
• Reproducibility. The helper scripts in each skill are version-controlled and documented. When I generate a plot using the plotting skill, the script that generated the plot is in my project. Six months later, I can re-run the script and get the same plot. The reproducibility is the part I appreciate most.

The reproducibility one is the part I underestimated. I used to generate plots interactively, save the final version, and lose the script. The library forces me to use a script, and the script is preserved. The discipline is the value.

What it is not good for

I want to be specific about the failure modes.

• The library is not a substitute for understanding the methods. The skills do the common operations. They do not teach the statistics. If you do not know what a t-test is, the t-test skill is not going to help you. The skill does the computation. It does not teach the concept.
• The library is not a model. It is a set of capabilities for an existing model. If the model is not good at the task, the library will not make it good. The library improves the model’s productivity on tasks it could already do. It does not give the model new abilities.
• The library is not production-grade. The skills are written by community contributors. The code quality varies. The documentation varies. The testing varies. For serious work, you are going to write your own version of the skills you use most. The library is a starting point, not a finished product.

The third one is the one I think most about. The library has been useful enough that I want to depend on it. The library has not been tested enough that I would bet my work on it. The bet is one I make consciously, knowing the library might break in a year.

What I would tell past me

If I could send a message back to the version of me that was about to use the library, I would say three things.

• Start with the data loading and plotting skills. Those are the most polished and the most useful for the first hour of looking at a new dataset.
• Read the SKILL.md file before you use a skill. The file tells you what the skill does, what it assumes, and what its failure modes are. The README is the documentation. The SKILL.md is the contract.
• Plan to rewrite the skills you use most. The library is a starting point. The skills you use every day are going to evolve into something specific to your workflow. That evolution is healthy. The library is the seed, not the tree.

Trade-offs

The library is tied to Claude. The skills are written for Claude’s context format and Claude’s tool-use system. If you switch to a different model, the skills do not transfer. The trade is “use the best model for the task” versus “use the model that is most portable.” The library chose the first.

Community-maintained, the library has responsive maintainers. The number of contributors is small. The skills are not going to disappear tomorrow. They are also not going to be polished at the rate of a corporate-backed project. The trade is community-driven evolution versus corporate-backed reliability. For a research tool, the first is fine.

Helper scripts are not always optimized. The plotting skill generates a publication-quality plot. It does so by running matplotlib with a specific style. The plot takes about 2 seconds to generate. The plot is not interactive. If you need an interactive plot, the skill is not the right tool.

Bottom line

claude-scientific-skills is the most useful scientific computing library I have used for an LLM in 2026. The skill structure is the right design. The library is not production-grade, and the skills are not a substitute for understanding. For exploratory data analysis, teaching, and reproducibility, the library is a real time saver. For production scientific computing, the library is a starting point.