Agent Skills

A collection of agent skills for our research group working on computational chemistry, materials science, and AI.

Each skill is a self-contained folder describing how an AI coding/research agent (e.g. Claude Code, but the format is intentionally generic) should carry out a recurring task — for example starting a new DFT calculation, submitting a job to the cluster, or setting up an ML training run.

You pick the skills you want and install them into your local agent's skills folder. Skills are independent: installing one does not require installing any other.

Existing skills

The skills currently in this repo:

Skill	Summary
xtb	Run the Grimme group's xtb semiempirical tight-binding program from the command line: geometry optimization, single-point energies, vibrational frequencies and thermochemistry, and implicit solvation (ALPB/GBSA) across the GFN0/1/2-xTB and GFN-FF methods — plus the newer general-purpose g-xTB method and CREST conformer/rotamer searches. The fast, cheap route to QM on a molecule, e.g. for pre-screening geometries before DFT.
chem-mat-database	Discover, download, and load chemistry & materials-science datasets from the ChemMatData database using the chem_mat_data package — both its cmdata CLI (list/info/stats/download) and its Python loaders. Pull any benchmark (clintox, esol, qm9, tox21, bace, lipophilicity, tmqm, …) as a pandas DataFrame (SMILES + targets), GNN-ready graph dicts, 3D structures, or transition-metal-complex tables, with PyG/Jraph conversion, streaming datasets for big data, and cache management. Pure data access — hands off to mol-gnn for training. Ships a self-contained uv script that tours the whole consumer API end-to-end.
mol-gnn	Train a graph neural network to predict molecular properties from a CSV of SMILES + targets, using chem_mat_data to featurize the molecules and PyTorch Geometric + Lightning for the model. Covers regression and classification, single- and multi-target, the model menu (GCN/GAT/GIN benchmarks vs. the recommended GIN/GINE/GATv2), train/val/test splitting, metrics, and gotchas — and ships a self-contained uv training script plus an example dataset that run end-to-end.
megan-xai	Train a MEGAN self-explaining GNN on a SMILES + target CSV: it predicts a property and produces per-node/edge attribution masks across multiple explanation channels. Drives the full loop — write a pycomex sub-experiment, smoke-test, run, read the automatic post-training diagnostic self-check, tune the explanation knobs when a run fails, and assemble a human-facing explanation report. Operates inside a graph_attention_student checkout.
md-multiatoms	Run molecular dynamics with ASE, scaled by the multiatoms package for batched, parallel dynamics on a GPU: replicate a structure into many systems and batch their force evaluations into one forward pass. Covers writing a ModelManager for any interatomic potential (a batched torch model, or an off-the-shelf ASE calculator like MACE/XTB/EMT), the MultiAtoms/PolyAtoms API, the setup-outside / step-inside parallel() rule, the minimize → equilibrate → produce workflow, and writing + analyzing trajectories (temperature, energy, RDF) — and ships a self-contained uv script that runs an EMT smoke test end-to-end.
autoslurm	Submit and monitor Slurm jobs with AutoSlurm (the aslurm CLI), which packs many commands into few jobs from a YAML template config. Covers config discovery, the sweep syntax (<[zipped]> vs. <{product}>), packing tasks onto GPUs, chain jobs for work longer than the walltime, and dry-running before submit — written mainly to pre-empt AutoSlurm's silent failure modes: output that looks like it went to /dev/null but didn't, a sweep that ran once because the angle brackets were unquoted or missing, and a config that is ignored because a same-named one shadows it.
pytorch-to-static-site	Assess feasibility and port PyTorch inference pipelines to a fully static browser site using ONNX export, client-side assets, ONNX Runtime Web, GitHub Pages, and GitHub Actions deployment.

Repository layout

Skills live as a flat list of folders at the repo root. Each folder is one skill:

agent-skills/
├── README.md                 ← you are here
├── CONTRIBUTING.md           ← how to write a new skill
├── _template-skill/          ← copy this to start a new skill
│   ├── SKILL.md              ← the instruction file (with frontmatter)
│   └── README.md             ← human-facing usage notes
├── xtb/                       ← run xtb / GFN-xTB / g-xTB / CREST calculations
│   ├── SKILL.md
│   ├── README.md
│   ├── references/           ← detail loaded on demand
│   └── assets/               ← example structure
├── chem-mat-database/         ← discover/download/load ChemMatData datasets
│   ├── SKILL.md
│   ├── README.md
│   ├── scripts/              ← runnable uv tour of the loader API
│   └── references/           ← catalog + graph-format docs
├── mol-gnn/                   ← train GNNs for molecular property prediction (PyG)
│   ├── SKILL.md
│   ├── README.md
│   ├── scripts/              ← runnable uv training script
│   └── assets/               ← example dataset
├── megan-xai/                 ← train self-explaining MEGAN GNNs + judge explanations
│   ├── SKILL.md
│   ├── README.md
│   ├── templates/           ← pycomex sub-experiment template
│   └── references/          ← diagnostic + knob docs loaded on demand
├── pytorch-to-static-site/    ← port PyTorch inference to static browser site (ONNX / WASM)
│   ├── SKILL.md
│   └── README.md
└── ...

Folders prefixed with _ (like _template-skill/) are not real skills — they are scaffolding/templates.

What a skill folder contains

File	Audience	Purpose
SKILL.md	agent	The skill itself: YAML frontmatter (name, description) + instructions.
README.md	human	Short description, when to use it, and install/usage notes.

Optional, add only if a skill needs them:

• scripts/ — helper scripts the skill invokes.
• assets/ — input templates, config files, reference data the skill copies or reads.
• references/ — longer docs the agent can load on demand.

See CONTRIBUTING.md for the full convention.

Installing a skill

There is no installer — you copy or symlink the skill folder into your agent's skills directory. For Claude Code that directory is ~/.claude/skills/.

Option A — copy (simple, frozen at install time):

cp -r start-dft-calculation ~/.claude/skills/

Option B — symlink (recommended; auto-updates when you git pull):

ln -s "$(pwd)/start-dft-calculation" ~/.claude/skills/start-dft-calculation

Repeat for each skill you want. To update copied skills, re-copy after pulling; symlinked skills update automatically.

To uninstall, remove the folder (or symlink) from ~/.claude/skills/.

Note: different agents look in different places. The skill content is portable; only the install location changes. Check your agent's docs for where it loads skills from.

Contributing

New skills are welcome and encouraged. Start by copying _template-skill/ and read CONTRIBUTING.md.