Scientific Agent Skills

Overview

Scientific Agent Skills is a library of scientific workflow instructions created by K-Dense for hosts that support the Agent Skills standard. The project README describes 177 skills spanning package use, database access, laboratory integrations, analysis and scientific communication. It is an instruction collection rather than a standalone runnable research agent or predictive model. Each skill uses a SKILL.md file to describe its purpose, workflow and version metadata, with examples, references, helpers or templates where applicable.

For chemistry-oriented work, the documented coverage includes RDKit and PyTDC guidance, molecular docking and ADMET workflows, calibrated 1D NMR processing, crystal structure analysis and computational chemistry. Related simulation guidance covers Cantera ignition-delay calculations, pycalphad phase equilibria and PyBaMM battery cycling. Database Lookup documents endpoint selection, pagination, access requirements and provenance for sources including PubChem and ChEMBL; these upstream databases remain separate resources, not services operated by the skill collection.

The collection supplies instructions and supporting assets to an agent host. Scientific inputs and outputs depend on the chosen workflow: examples described in the README include spectra, crystal structures and simulation inputs, leading to processed spectra, structure analyses or simulation results. Dependencies, credentials and network access must be arranged separately according to each skill. Host discovery and optional metadata handling vary, and local tests do not establish live-service reliability or scientific validity. Skills can direct code execution and file changes, so users should inspect a focused subset before installation and evaluate results in their own environment.

Key Features

• Per-skill `SKILL.md` documentation records purpose, workflow and version metadata, with supporting examples, references, helpers or templates depending on the skill.
• Chemistry guidance covers RDKit, PyTDC, molecular docking, ADMET analysis and calibrated 1D NMR processing.
• Materials and engineering workflows include crystal structure analysis, pycalphad phase equilibria, Cantera ignition delay and PyBaMM battery cycling with documented numerical checks.
• Database Lookup documents 80 sources, including PubChem and ChEMBL, with endpoint selection, pagination, access requirements and provenance checks.
• Supports Agent Skills packaging and an Agent Plugins layout comprising `plugin.json` and `skills/`, with installation and discovery behavior dependent on the host.
• Skills with bundled `scripts/` have corresponding test suites and dependency entries; the README distinguishes structural checks from scientific-dependency testing.

Use Cases

• Intended evaluation: assess whether RDKit or PyTDC guidance helps an agent carry out a molecular-analysis workflow on representative inputs, checking generated results independently.
• Intended evaluation: use the documented database-retrieval procedures to assemble provenance-tracked PubChem or ChEMBL records for a chemistry research task.
• Intended evaluation: compare the calibrated 1D NMR workflow against a trusted spectrum-processing procedure before applying it to research data.
• Intended evaluation: assess the Cantera, pycalphad or PyBaMM guidance for a defined simulation problem, examining input provenance and numerical checks.

How to Use

1. Read the official README and select a small set of skills relevant to your task, such as RDKit, nmrglue or pycalphad. Avoid installing unrelated workflows initially.
2. Inspect the selected SKILL.md files and supporting assets in the repository. Check external connections, file operations, compatibility requirements and individual skill licenses before trusting them.
3. Choose a documented installation route for your host. Consult the Agent Skills standard and your host documentation to confirm discovery paths and optional metadata behavior; record the repository revision used.
4. Prepare the dependencies, reference data and credentials required by those skills. Keep incompatible workflows in separate environments; repository tooling requirements are not universal requirements for every scientific package.
5. Confirm that the host discovers the selected skills, then request a bounded workflow using representative inputs. Evaluate outputs against trusted references, retaining input provenance and validation records before expanding use.