The DIGIT-X lab's shared collection of Claude Code skills.
ATELIER is where the lab keeps reusable, version-controlled skills — packaged instructions and tools that make Claude Code do specialised research work the same high-quality way every time. Install it once and every lab member gets the same capabilities; improve a skill and everyone benefits.
Research work is full of high-stakes tasks that are easy to do adequately and hard to do well and consistently — a publication figure, a concept diagram, a grant proposal, a peer review. Done ad hoc, every person (and every paper, every application) comes out different and quality drifts. ATELIER captures the lab's way of doing these things as skills, so the output is reproducible, on-brand, and submission-grade by default — not dependent on who happened to make it.
It started with research visuals and is growing into a broader research toolkit — for both building research writing (visuals, grant proposals) and stress-testing it (reviewing papers and theses) — anything worth codifying so the whole lab does it the right way every time.
A skill is a folder with a SKILL.md (instructions Claude follows) plus any supporting tools/assets.
Claude loads a skill only when it's relevant. Skills are shareable and composable — this repo is a
Claude Code plugin marketplace bundling the
lab's skills.
| Skill | Status | What it does |
|---|---|---|
research-visuals |
✅ available | A design assistant for reproducible, journal-grade visuals. You choose a colour palette (from Sanzo Wada's A Dictionary of Color Combinations) and fonts once; it locks them to a theme.json and every figure, diagram, and slide inherits that look. Four families: figures (D3.js → SVG/PDF, never matplotlib), imaging (niivue NIfTI/DICOM renders), concepts (HTML5 Canvas + Phosphor icons), and slides (reveal.js). It recommends the right chart for your data, previews live in the browser, and runs an automated craft gate so nothing ships cluttered or overlapping. |
grant-proposals |
✅ available | A grant-writing assistant for DFG · ERC/Horizon · NIH · industry funders. It works from your real science (it never fabricates) — interviews you, ingests your call/RFA + papers, grounds every claim in real literature via Paperclip, pressure-tests the idea before you write a word (is it actually novel/fundable?), drafts to the agency's best practices, machine-checks limits + citations, and writes it into Google Docs with figures + a Gantt from research-visuals. Concierge setup walks anyone through it. |
academic-review |
✅ available | A reviewing assistant for theses, papers, and any academic work — the critique/improve counterpart to grant-proposals. Two inferred modes (get tough pre-submission feedback on your draft, or produce a submittable peer-review referee report on someone else's), four review lenses (rigor · clarity · integrity · literature), reporting-guideline packs (CLAIM/TRIPOD-AI/STROBE/CONSORT/STARD/PRISMA/ARRIVE), an opt-in adversarial deep-panel, and a machine gate that holds the review itself to a located-and-evidence-anchored bar. |
As a plugin (recommended for the lab):
/plugin marketplace add DIGIT-X-Lab/atelier
/plugin install atelier
Standalone (symlink into your personal skills):
git clone https://github.com/DIGIT-X-Lab/atelier.git
bash atelier/install.sh
- Reproducible by default — choose the look once, lock it, reuse it everywhere.
- Submission-grade — vector output, journal/grant-aware sizing and fonts, accessibility checks.
- Ask, don't assume — e.g. fonts are always chosen by you, never silently defaulted.
- A work of art, never cluttered — every visual passes a no-overlap / breathing-room craft gate.
.claude-plugin/marketplace.json # makes this repo /plugin-installable
plugins/atelier/ # plugin manifest (skills symlinked from ./skills)
skills/<skill-name>/SKILL.md # the skills themselves (single source of truth)
install.sh # standalone symlink installer
SPEC.md # design spec for research-visuals
CONTRIBUTING.md # how to add a skill or a chart recipe
See CONTRIBUTING.md. In short: add skills/<name>/SKILL.md (it's picked up
automatically), keep SKILL.md lean with heavy reference in sibling files, and test it on a real task
before relying on it.
Maintained by the DIGIT-X Lab · LMU Munich.