Ccf Figuresmith Skills — Data skill for Claude Code

ccf-figuresmith-skills

A Claude Code skill for producing the figures that go into a real paper submission. Built and debugged on an actual TRB submission, so every rule in it traces to a failure that reached a draft at least once.

 

*Both produced by the bundled templates: `assets/figstyle.py` and `assets/build_drawio_skeleton.py`, each run as its own smoke test.*

What it covers

Both figure populations a paper has, which need completely different tools but the same discipline:

• Data figures (Results): matplotlib, driven by a results file, laid out at the true printed size and gated by a bounding-box QA check that measures every label's rendered box, every panel's alignment, and the palette's actual colour-vision separation.
• Conceptual diagrams (Intro / Method): a script that emits draw.io XML, gated by a text-fit check, with an icon set and helpers for grouping, branching, and feedback loops.

Deliverable for both is a **PNG** — 400 dpi from matplotlib, or exported from an editable `.drawio` source. Vector output is one flag away when a venue demands it.

The shared discipline is what makes it work: generate rather than hand-make, lay out at the true printed size, gate automatically, **look at the rendered raster**, iterate more than once, then verify on the compiled page.

It also covers the part before any of that: `references/results-repertoire.md` is about enumerating how many different true things one results file can say, because a results section with three near-identical bar charts usually failed at that step rather than at the drawing.

**Delivery branches inside the skill, and the default is local.** It checks the git remote. For an ordinary local or GitHub repo it finishes at a local commit and does not push. If it finds `git.overleaf.com` it says so and asks, defaulting to not publishing, and only then reads `references/overleaf.md`. A project wi