analytical-figures-skill

analytical-figures

Ask Claude Code for a calibration curve from your .spc files and get a journal-ready figure with LOD/LOQ, confidence and prediction bands and a residual panel, plus one self-contained Python script that reproduces it. Ask for a calculated PXRD pattern from a CIF, a cocrystal-vs-physical-mixture call or PLS diagnostics with leakage-safe cross-validation and get the same: a figure that passed its checks, and the code that made it.

analytical-figures is a Claude Code skill for analytical chemistry and spectroscopy: FTIR/ATR and PXRD spectra, calibration with figures of merit (LOD/LOQ, recovery, ICH Q2), multivariate calibration (PLS/PCR/PCA), crystal structures from CIFs (validation table, calculated PXRD, deterministic 3D views) and cocrystal identification.

It is not a plotting tutorial. It does two things a plotting library does not:

• It gates the numbers before they reach a figure. Ingest checks, identical processing across a batch, named error statistics (SD vs SEM vs CI, with n), leakage-safe cross-validation, mandatory residual panels, no extrapolation past the calibrated range, and a static critic that flags a figure-of-merit hand-typed onto a figure instead of interpolated from the code.
• It hands over a script, not a PNG. The deliverable is one self-contained Python file (bundle.py amalgamates the modules an analysis uses) with a provenance header, so every figure is reproducible, editable and attachable to a report.

Hover a tile for what it is and where the data came from; click for full size. Every tile is produced by docs/build_gallery.py from public data and public CIFs — nothing is drawn by hand and nothing is synthetic. Interactive version with placards: https://thedop.github.io/analytical-figures-skill/

From a prompt to a figure

Three prompts, run in this repository with the skill loaded. Left: the prompt and the gates that fired on the way. Right: what came back. Every run hands over one self-contained script that regenerates its figure, and the caption is interpolated from the computed values, never typed.

Install

As a plugin, from inside Claude Code (two commands; the repo is its own marketplace):

/plugin marketplace add TheDop/analytical-figures-skill
/plugin install analytical-figures@analytical-figures-marketplace

Or clone into a Claude Code skills directory, user-level or per project:

git clone https://github.com/TheDop/analytical-figures-skill ~/.claude/skills/analytical-figures        # available in every project
git clone https://github.com/TheDop/analytical-figures-skill .claude/skills/analytical-figures           # this project only

Either way, install the Python side once:

python -m pip install -r requirements.txt        # from the cloned/installed skill directory

numpy and matplotlib are required; scipy is strongly recommended. Everything else is per-family and imported lazily, so an FTIR job never pulls it: requirements.txt adds scikit-learn (chemometrics) and rdkit (the cocrystal predictor); the crystal family needs python -m pip install gemmi Dans-Diffraction (plus pymatgen for the PXRD cross-check and pyvista for the VTK renders), which the skill asks for by name the first time a CIF job runs. Claude Code loads the skill at session start. Then ask for what you need: a replicate overlay with %RSD, a calibration curve with LOD/LOQ, a calculated PXRD pattern from a CIF, a cocrystal-vs-mixture call.

What is inside

The modules are plain Python, so the skill also works without Claude: run from the skill root and from scripts import spectra, calibration as the template does.

Validation

python -m pytest tests/ is one green gate for the whole skill (81 tests). It checks module imports, the SKILL.md structure, the bundle.py round-trip, p-value correction against statsmodels, the style presets and the CLI, and it bridges in three numeric-parity suites:

| skill_validation/ftir_integration/ | 14 | closed-form Gaussian areas and scipy.integrate.quad under sloping baselines, noise and band overlap; plus the integrator run against the published pectin-DM (Wang 2023) and BMC 2021 calibrations | | skill_validation/crystal/ | 9 CIFs | a CIF edge-case set (disorder groups, special positions, Z′ = 9, chiral, multi-block, synthetic broken inputs): the engine's table, an independent gemmi-only second opinion, ADP U_eq, the Dans reflection list vs pymatgen. Needs gemmi + Dans_Diffraction; pytest skips it otherwise |

The 24 pectin spectra (CC BY 4.0) are vendored. Data that cannot be redistributed is fetched on demand instead: the biospectools EMSC gold and the R pls gasoline set (skill_validation/chemometrics/datasets/fetch.py) and the CCDC CIFs. Every check that needs one of them skips cleanly when it is absent.

Known limitations

• The crystal family is written for small-molecule cells. Bond, geometry and H-bond searches use a KD-tree on a cached supercell, but the first calculated pattern per structure still parses the CIF and computes structure factors (seconds for a few hundred atoms); later calls reuse the cache.
• The realistic PXRD profile evaluates each reflection within ±40 FWHM of its centre (window_fwhm); the discarded Lorentzian tail is below counting quantisation on lab patterns, and window_fwhm=None restores the exact full-grid sum.
• audit_layout is a set of heuristics (tick overlap, clipping, letters, unit cues). It catches the common faults; reading the saved PNG is still part of the workflow, not a formality.

Provenance and licence

Built during an MSc analytical-chemistry project: quantifying an API in a binary excipient mixture by ATR-FTIR, then identifying cocrystals by FTIR and PXRD. The worked examples in the docs (aspirin in lactose, theophylline cocrystals) come from that work.

MIT licence (see LICENSE). The figure-selection scaffolding and the panel-label alignment trick are adapted from scipilot-figure-skill (MIT, Haojae); details in THIRD_PARTY_NOTICES.md.