| name | viz-figure-mirror |
| description | Transfer the visual style of a reference paper figure onto the user's own data. Takes a screenshot of a figure from a paper plus tabular data and produces an editable matplotlib script in that figure's register, gated by a mechanical quality floor and a Drawer/Reviewer refinement loop. Output is a self-contained .py + PNG + Type-42 PDF. Triggers on: "match this figure style", "figure in the style of", "make my plot look like this paper", "publication figure from reference", "style transfer figure", "reproduce this figure with my data", "NeurIPS-quality version of this chart". Does NOT trigger for: plain data plots with no reference image (use sci-data-analysis), AI-generated illustrations (use viz-nano-banana), flowcharts/architecture/sequence diagrams (use viz-diagram-code), hand-drawn sketches (use viz-excalidraw-diagram).
|
Figure Mirror
Render the user's data in the visual style of a reference figure from a top-tier
paper. The output is not a copy of the reference; it is the user's numbers wearing
the reference's typographic voice, palette, spine treatment, gridline weight, and
layout density. The deliverable is an editable matplotlib script the user owns,
plus a rendered PNG and a camera-ready Type-42 PDF.
Inspired by FigMirror (open project, no license file); the floor algorithm,
prompt prose, and convention values here are an independent clean-room
reimplementation.
Outcome
A workspace under projects/viz-figure-mirror/{slug}/:
figure.py โ self-contained script with an inline DATA SECTOR and
pdf.fonttype = 42. The user edits this.figure.png โ rendered preview.figure.pdf โ Type-42 camera-ready export.- per-iteration artifacts (
figure_iter*.py, img_iter*.png, audit_iter*.json).
Input / output contract
| In | Out |
|---|
| Reference image (PNG/JPG screenshot of a paper figure) | figure.py (editable matplotlib) |
| User data (CSV / TSV / pasted table / dirty terminal text) | figure.png + Type-42 figure.pdf |
Trigger requires both a reference image and data. With only data and no
reference, route to sci-data-analysis instead.
Context Needs
| File | Load level | How it shapes this skill |
|---|
research_context/research-profile.md | โ | Not used |
context/learnings.md | ## viz-figure-mirror section | Apply previous feedback before starting |
The two-role loop
The work is split between two roles that never talk directly. The orchestrator
(this session) shuttles artifacts between them.
- Drawer (
references/drawer.md) โ writes figure_iter<N>.py, renders it,
runs its OWN floor self-check (scripts/figure_quality.py) before handoff, and
records anchor measurements at iter 0. Grounds every choice in L1 or L2.
- Reviewer (
references/reviewer.md) โ a fresh-context vision-only audit. Sees
the reference + the draft + the L2 library + the prior audit (never the data, never
the code). Emits ONE strict JSON object validated by scripts/review_schema.py.
Full wiring, stop conditions, and the select-best fallback live in
references/orchestration.md.
The grounding hierarchy (L1 / L2 / L3)
Every property of the figure traces to exactly one source:
- L1 โ the reference image. Highest authority. What the reference visibly does
is what the draft does. Measure PIL-reliable properties (full-image aspect,
palette of large filled regions); eyeball the brittle ones with acknowledged
uncertainty.
- L2 โ
references/aesthetic-library.md. Paper-figure conventions. The fallback
class vocabulary for properties PIL cannot measure reliably (spine colour/width,
gridline width, font weight) and the extension menu when the data has more series
than the reference.
- L3 โ your own taste. Banned. "I think it would look better" diverges the
loop. If a choice cannot cite L1 or L2, drop it.
The quality floor (mechanical, deterministic)
Before any aesthetic judgement, a figure must clear a mechanical floor enforced by
scripts/figure_quality.py. The floor is binary and catches the two defect classes
that make a figure unshippable regardless of style:
assert_no_text_overlap(fig) โ no two visible text bboxes (tick labels, titles,
axis labels, annotations) may overlap.assert_no_clipped_labels(fig) โ no text bbox may fall outside the canvas (tick
labels centred on an axis extreme get a small font-proportional allowance; titles,
axis labels, and annotations stay strict).assert_no_legend_data_overlap(fig) โ no legend may sit on top of plotted data ink
(bars, markers, lines, error bars). Conservative: a legend over empty whitespace, or
one placed outside the axes, passes; only a legend clearly occluding data is flagged.check_floor(fig) -> {"passed": bool, "violations": [...]} โ same checks, no
raise, for the loop (violation kind โ text_overlap / label_clipped /
legend_over_data).
The module also carries PUBLICATION_RCPARAMS (anti-slop: hairline left+bottom
spines, no tick marks, light-grey gridlines, Type-42 fonts, real minus glyph) and
apply_publication_style(ax). The Drawer self-checks the floor before every
handoff; the Reviewer never sees a draft that fails the Drawer's own check.
Workflow
- Stage workspace. Create
projects/viz-figure-mirror/{slug}/inputs/. Save the
reference to reference.png, the parsed data to data.txt.
- Echo the data parse. Show the user rows x cols, columns, NaN cells, a sample
row. Confirm before drawing.
- Iterate up to
max_iters (default 6). Per iter: Drawer renders + self-checks
floor -> Reviewer audits -> validate JSON -> decide.
- Stop when
floor.passed && verdict == "ship". Otherwise at the budget, run
the select-best fallback (select_best in scripts/review_schema.py).
- Write canonical artifacts (
figure.py / .png / Type-42 .pdf), then offer
the Drive Push Gate and Figure Proposal follow-ups per CLAUDE.md.
Self-containment
This skill carries its OWN floor module (scripts/figure_quality.py) and its OWN
schema validator (scripts/review_schema.py). It does not import from other skills.
Dependencies
| Dependency | Required? | Provides | Fallback without it |
|---|
matplotlib | Required | Figure rendering, the mechanical quality floor, Type-42 PDF export | None โ the skill is non-functional |
pillow (PIL) | Required | PIL-measurement of the reference image (aspect, palette of large filled regions) | None โ L1 measurement + floor color checks non-functional |
numpy | Required (transitive via matplotlib) | Array handling in generated scripts | None โ installed with matplotlib |
| Vision-capable model | Required for the Reviewer audit | Multimodal reference-vs-draft critique that drives the refinement loop | Floor-only mode โ the loop runs the mechanical floor + select_best without aesthetic review (see references/orchestration.md โ "Reviewer unavailable") |
Run bash scripts/setup.sh once per machine to verify/install matplotlib + pillow.
No API key and no network are needed for the mechanical floor; only the Reviewer audit
needs a vision-capable model.
