Agent Skills

Agent skills are reusable instruction packages that extend the agent’s capabilities for specific tasks. They live in ~/.agents/skills/ on your machine and are loaded on demand.

Full catalog and installation instructions: → K-Dense-AI/scientific-agent-skills

Bundled simulation skills (shipped with this template)

Three simulation launch skills are included in .github/skills/ and are used automatically by @simulation-agent. Each ships a Pydantic-validated Python runner that patches configuration files, launches the simulation, and returns a SUCCESS / ERROR status line.

Skill Simulation code Use case
dustpy DustPy Radial dust evolution, grain growth, fragmentation barrier, Stokes numbers, dust-to-gas mass fractions
fargo3d FARGO3D Planet–disk interaction, gap opening, type-I/II migration torques — patches .par file without recompiling
pluto PLUTO HD/MHD disk and jet simulations — overrides pluto.ini parameters without recompiling

Each skill follows the lean SKILL.md pattern: the SKILL.md file stays compact (trigger conditions, procedure, key parameters, output format, error table). Full parameter tables and worked examples live in the skill’s references/ subdirectory and are read by the agent on demand:

skills/<code>/
├── SKILL.md              # Lean — quick-scan essentials only
├── references/
│   ├── parameters.md     # Full parameter table (types, defaults, constraints)
│   └── examples.md       # Step-by-step run examples  (PLUTO only)
└── scripts/
    └── run_<code>.py     # Validated runner

Prerequisites:

  • dustpy: pip install dustpy scientific-pydantic
  • fargo3d: compiled binary (fargo3d) must already exist — the skill patches .par files and launches an existing binary; recompilation (e.g. when changing NFLUIDS, MHD, or setup directory) must be done manually with make in the FARGO3D root
  • pluto: no pre-existing binary required — compile_pluto.py builds or rebuilds ./pluto automatically when physics, geometry, dimensions, EOS, or module flags change; the skill’s pre-flight checklist (STEP 0) determines whether recompilation is needed before each run
# Example: launch DustPy via simulation-agent
@simulation-agent  Run a dust evolution simulation with alpha=1e-3,
                   disk mass 0.05 Msun, fragmentation velocity 10 m/s,
                   for 1 Myr. Save snapshots to data/dustpy/run01/.

Universal (all groups)

Skill Best used for
astropy Coordinate transforms, FITS I/O, cosmological distances, WCS, time systems
matplotlib Publication plots requiring fine-grained control over every element
scientific-visualization Multi-panel journal figures (Nature/A&A style, colourblind-safe palettes, significance annotations)
statistical-analysis Choosing appropriate tests, assumption checking, APA-formatted results
paper-lookup Searching PubMed, arXiv, OpenAlex, Semantic Scholar, Crossref
citation-management Verifying BibTeX, DOI → BibTeX conversion, reference accuracy

Disk & planet formation

Skill Best used for
database-lookup Querying SIMBAD, VizieR, ALMA archive, ExoFOP, Gaia DR3
exploratory-data-analysis First look at a new simulation output or observational data file
scientific-schematics Disk structure diagrams, gap morphology schematics, protoplanetary disk cross-sections

Cosmological simulations

Skill Best used for
networkx Merger trees, substructure graphs, galaxy filament networks
umap-learn Dimensionality reduction for halo/galaxy property distributions
scikit-learn Classification / regression on simulation catalogues
shap Interpreting ML models trained on halo catalogues

Atmospheric retrievals & high-res spectroscopy

Skill Best used for
statsmodels Frequentist inference, time-series detrending
shap Interpreting ML-based retrieval or classification models
database-lookup Querying ExoAtmospheres, HITRAN, ExoMol, NASA Exoplanet Archive
aeon Time-series classification of stellar / planetary light curves

X-ray & galaxy clusters

Skill Best used for
scikit-survival Survival / time-to-event modelling (e.g. cluster cooling time distributions)
scientific-visualization Thermodynamic maps, surface brightness profiles

How to use a skill

In Copilot Chat or Agent Mode, tell the agent which skill to load at the start of a task:

Use the scientific-visualization skill for this figure.
Plot the dust surface density profile with Nature-journal styling and
save to plots/disk_sigma.pdf.

The agent reads the SKILL.md file and follows its step-by-step instructions. Skills can be chained — e.g. paper-lookup to find references, then citation-management to produce clean BibTeX.