An original, runnable Codex skill suite for turning raw research data and scientific images into publication-ready figures, then auditing the final artwork.
中文简介：从原始表格自动生成多种可靠的 SCI 图候选并一键换色；批量统一显微、荧光、电镜等科研图片的尺寸、显示参数和规范标尺；最后完成可编辑 SVG、固定画布、字体、留白、重叠与组图质控。
These are scientific figures, not promotional artwork. Every observation, state transition, network weight, matrix value, and model result is deterministic synthetic data generated by the repository. Captions stay outside the reusable artwork.
The same simulated experiment is expressed as a conserved cell-state alluvial map, directional ligand-receptor network, RNA-ATAC concordance map, and treatment-response ridgelines. Flow totals are checked in code before export.
Bent, non-isotropic single-cell manifolds, pseudotime branching, spatial domains with a declared synthetic scale, and a marker dot plot are combined without evenly spaced circular islands or generic geometric tissue placeholders.
Directional cell communication, cross-platform pathway activity, a causal mediation graph, and an optimization landscape share one visual system while keeping their scientific encodings distinct.
Feature contributions, calibration, decision-curve analysis, and repeated external validation show model behavior beyond a single headline AUC.
Reproduce all showcase figures and editable SVG/PDF masters from a fresh clone:
python -m pip install -r requirements.txt
python demo/figure_sources/make_demo_suite.py| Skill | What it does | Main outputs |
|---|---|---|
make-sci-data-figures |
Profiles CSV/TSV/XLSX data, records the experimental design, chooses defensible statistics, and creates several chart candidates from the same data | PNG/SVG/PDF candidates, gallery, analysis plan, reproducible recipe |
standardize-sci-images |
Non-destructively standardizes microscopy, fluorescence, histology, and electron-microscopy images | Equal-size panels, calibrated scale bars, montage, SHA-256 processing audit |
polish-sci-figures |
Redraws, assembles, and audits final manuscript, slide, poster, or showcase figures | Fixed-canvas editable files and final-size QA |
These are not renamed copies of third-party skills. The workflow and code were written for this repository around recurring real-world pain points: wrong statistical units, hidden distributions, inconsistent palettes, unequal canvases, changing apparent font sizes, internal mini-titles, panel numbers, overlaps, broken scientific notation, uneditable SVG text, guessed scale bars, and unfair per-image contrast tuning.
Give it a tidy table, name the group, outcome, biological experimental unit, and whether observations are independent or paired. It profiles the data, matches a defensible analysis, generates several genuinely different figure types, and preserves the decision in a machine-readable analysis plan. The result is not a decorative chart template: raw observations, effect size, uncertainty, design, exclusions, sensitivity analysis, canvas, typography, and palette semantics stay connected.
- High-information alternatives: estimation graphics, rainclouds, paired estimation graphics, group-estimate intervals, raw-data summaries, box plots, and violins are generated only when the design and sample size support them.
- Statistics before decoration: the biological unit and pairing determine the analysis; the preferred visual style cannot silently change the test.
- One-command restyling: palette changes leave data, statistics, axes, labels, ordering, and geometry untouched.
- Manuscript-ready construction: identical physical canvases, readable Arial text, no internal mini-titles or panel numbers, live SVG text, and editable PDF/SVG output.
- Refuses false certainty: unsupported repeated measures, mixed models, survival analysis, count models, compositional data, and high-dimensional omics are routed to specialist analysis instead of receiving an invented automatic test.
All values below are deterministic synthetic demonstration data. Descriptions remain outside the reusable artwork.
Raw observations and group estimates are shown beside the bootstrap distribution of the treatment-minus-control effect. In the bundled example the estimated mean difference is 2.48 a.u. (95% CI 1.95 to 3.00), so the figure communicates magnitude and uncertainty rather than relying on a significance star.
Half-violin density, every observation, interquartile range, and median are combined without hiding the sample. The workbench does not generate this density view when a group has fewer than 10 observations.
Subject-level trajectories preserve the matching, while the second axis shows the within-subject mean difference and 95% CI. In the example the paired change is 1.14 normalized units (95% CI 0.978 to 1.31).
Every biological sample remains visible behind the group mean and 95% CI. The global analysis is kept global; the skill does not manufacture pairwise claims that were never specified.
These choices follow estimation-first reporting described in Ho et al., Nature Methods (2019), the transparent distribution logic of raincloud plots, and the editable-text, accessible-color, compact-layout guidance in the Nature Research figure guide. A good figure cannot rescue a weak experimental design or guarantee acceptance, but it can remove avoidable statistical and artwork failures.
| Research situation | Minimal user input | Automatic match and concrete example | Scientific guardrail |
|---|---|---|---|
| Control vs treatment using different biological samples | condition, Response, sample_id, independent |
Treatment-minus-control mean difference and 95% CI; Welch two-sample test; Mann-Whitney sensitivity analysis. Example effect: 2.48 a.u. (1.95 to 3.00). | Repeated unit IDs are rejected instead of being counted as independent replication. |
| Before vs after on the same subjects | condition, Response, subject_id, paired |
Within-subject mean difference and 95% CI; paired test; Wilcoxon sensitivity analysis. Example change: 1.14 normalized units (0.978 to 1.31). | Duplicate subject-condition rows are rejected; incomplete pairs are counted and reported. |
| Vehicle plus three independent dose groups | condition, Response, sample_id, independent |
Group means and 95% CIs; global Welch ANOVA; Kruskal-Wallis sensitivity analysis. | A global result never becomes an undeclared pairwise claim; multiplicity remains explicit. |
Exact tests, effect direction, sample counts, exclusions, diagnostics, limitations, analysis scope, and multiplicity status are written to analysis_plan.json. Exploratory results show the effect and interval on the artwork while keeping P values in the analysis record; --scope confirmatory --show-p-value is available only after a pre-specified confirmatory family is declared.
# Independent two-group comparison: creates five candidates
python skills/make-sci-data-figures/scripts/figure_workbench.py generate \
skills/make-sci-data-figures/examples/synthetic_group_comparison.csv \
--group condition --value Response --unit sample_id \
--design independent --order Control,Treatment --unit-label "a.u." \
--outdir demo/workbench
# Paired before/after comparison: creates paired effect and trajectory views
python skills/make-sci-data-figures/scripts/figure_workbench.py generate \
skills/make-sci-data-figures/examples/synthetic_paired_response.csv \
--group condition --value Response --unit subject_id \
--design paired --order Before,After --unit-label "normalized units" \
--outdir demo/paired_workbench
# Four independent groups: creates group-interval and distribution views
python skills/make-sci-data-figures/scripts/figure_workbench.py generate \
skills/make-sci-data-figures/examples/synthetic_multigroup_response.csv \
--group condition --value Response --unit sample_id --design independent \
--order "Vehicle,Low dose,Mid dose,High dose" --unit-label "a.u." \
--outdir demo/multigroup_workbenchChange only the palette while keeping the analysis and geometry fixed:
python skills/make-sci-data-figures/scripts/figure_workbench.py recolor \
demo/workbench/figure_recipe.json --palette okabe_ito \
--outdir demo/workbench_okabe_ito
The current automatic inference scope is deliberately limited to common continuous-outcome independent and paired group comparisons. Unsupported specialist designs receive an explicit limitation instead of a plausible-looking but scientifically unsafe answer.
The preview below uses synthetic fluorescence-like software-test images, not biological observations.
python skills/standardize-sci-images/scripts/make_example_data.py \
--outdir demo/image_inputs
python skills/standardize-sci-images/scripts/standardize_images.py \
demo/image_inputs/manifest.csv --scale-bar-um 20 \
--outdir demo/image_standardizationThe image workflow never overwrites raw files, never invents calibration, never tunes display settings per comparison image, and never resamples by default. It records source hashes, crop boxes, display parameters, calibration, and scale-bar geometry. Each panel includes an unannotated display raster, a review preview, and an SVG with the scale bar/text kept as editable vector/live layers.
All values in these previews are deterministic synthetic demonstration data.
- No panel letters, serial labels, internal titles, or subtitles unless the verified target explicitly requires them.
- Arial by default, with one-place switching to Times New Roman or another verified journal font.
- Correct scientific case, italics, symbols, units, subscripts, and superscripts.
- Zero unintended overlap at final placement.
- Equal physical canvases and axes geometry for panels that will be assembled together; no tight-crop export.
- Editable SVG/PDF plus high-resolution PNG, with live continuous text.
- Stable group order, color meaning, uncertainty definition, and statistical scope.
- Raw scientific data and images remain authoritative; examples stay clearly labeled synthetic.
Clone or download the repository, install dependencies, then copy all three skill folders.
python -m pip install -r requirements.txt
New-Item -ItemType Directory -Force "$HOME\.codex\skills" | Out-Null
Copy-Item -Recurse -Force ".\skills\make-sci-data-figures" "$HOME\.codex\skills\"
Copy-Item -Recurse -Force ".\skills\standardize-sci-images" "$HOME\.codex\skills\"
Copy-Item -Recurse -Force ".\skills\polish-sci-figures" "$HOME\.codex\skills\"python -m pip install -r requirements.txt
mkdir -p ~/.codex/skills
cp -R skills/make-sci-data-figures ~/.codex/skills/
cp -R skills/standardize-sci-images ~/.codex/skills/
cp -R skills/polish-sci-figures ~/.codex/skills/Start a new Codex session after installation.
Use $make-sci-data-figures to profile this table and make several publication-ready candidates.
Use $make-sci-data-figures to rerender the selected figure with the okabe_ito palette only.
Use $standardize-sci-images to standardize this microscopy batch and add calibrated 20 µm scale bars.
Use $polish-sci-figures to assemble the chosen panels and audit the final editable SVGs.
python skills/make-sci-data-figures/scripts/test_figure_workbench.py
python skills/standardize-sci-images/scripts/test_standardize_images.py
python -m compileall -q demo skillsFor a set of independently editable SVG panels intended for the same slot:
python skills/polish-sci-figures/scripts/check_svg_canvas.py path/to/panels/*.svg
python skills/polish-sci-figures/scripts/check_svg_editability.py --require-fully-editable path/to/panels/*.svgskills/make-sci-data-figures/ raw data, statistics, candidate charts, palette recipes
skills/standardize-sci-images/ calibrated image standardization and processing audit
skills/polish-sci-figures/ final drawing, assembly, export, and QA
demo/ reproducible synthetic previews
requirements.txt Python dependencies
MIT License. See LICENSE.