Turn experimental data into publication-ready scientific figures — directly inside Claude Code.
English · 中文 · Gallery · Workflow · Quick Start
<a id="english"></a>
SciFig Generate is a Claude Code skill that transforms experimental data — CSV, TSV, Excel, or matrix files — into submission-ready scientific figures.
Drop the skill into Claude Code, mention your data file, and let the four-phase pipeline handle ingestion, chart planning, statistical-test selection, journal-styled code generation, and export.
The skill is self-contained: 121-chart registry, six journal-style profiles (Nature / Cell / Science / Lancet / NEJM / JAMA), zero-touch finalizer (legend contract, layout audit, heatmap label sizing, text-occlusion guards), and SVG / PDF / PNG export with source data, render-QA evidence, and methods-ready statistical reports. No pip install required.
- V0.3.0: Skill-only repository — removed pip package (
src/), release scripts (scripts/), tests, and PyPI/CLI docs. All runtime behavior lives inscifig/. - V0.2.1: Refactored the scifig skill — split the 8341-line
generators-distribution.mdmonolith into 13 legal domain-splitgenerators_<domain>.pymodules; slimmedcase-index.jsonby moving bulk evidence into gitignoredcase-evidence.json; Phase 3 now injects generator source on-demand via_build_generator_code. - V0.1.2: Add editable SVG output with matching PNG regeneration.
- V0.1.1: Fix gallery style consistency for centered titles, compact bottom legends, outside panel labels, and ASCII-safe scientific labels.
- V0.1.0: Initial release.
<a id="gallery"></a>
All figures below were generated by SciFig from real open-source datasets across 6 scientific domains. No manual post-processing.
<a id="workflow"></a>
SciFig runs as a Claude Code skill. The entire process is triggered inside a Claude Code conversation — no separate server, no API key, no browser tab.
User types trigger keyword or /skill scifig
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ SKILL.md — Coordinator │
│ Validates file path, reads data, dispatches phases │
└───────────────────────────┬─────────────────────────────────────┘
│
┌────────────────────┼────────────────────┐
▼ ▼ ▼
Phase 1 Phase 2 Phase 3
Data Detection Chart Planning Code Generation
────────────── ──────────── ────────────────
• Validate file • Recommend chart • Apply journal
path + encoding families by kernel (font,
• Infer structure domain + data size, layout,
(tidy/wide/ shape dpi)
matrix) • Plan panel • Generate
• Detect domain blueprint matplotlib code
(13 scientific • Select • Run finalizer
domains) statistical (3 auto-correct
• Build dataProfile tests passes)
• Collect • Build palette • Layout audit
preferences plan • Legend contract
• Return: • Return: • Return:
dataProfile chartPlan styledCode
│ │ │
└────────────────────┼────────────────────┘
▼
Phase 4
Export & Report
────────────────
• Export SVG/PDF
• Generate source data tables
• Render QA evidence
• Write methods-ready stats report
• Output reproducible code snapshot
• Return: outputBundle
<a id="quick-start"></a>
Step 1 — Install
git clone https://github.com/Techd81/SciFig.git
cp -r SciFig/scifig ~/.claude/skills/Restart Claude Code. The skill is auto-discovered.
Step 2 — Trigger
In any Claude Code conversation, mention your data file. Trigger keywords: generate figure, plot data, sci figure, 科研图, 画图, 多 panel. Or use the explicit command:
> /skill scifig
> FILE: /path/to/your_data.csv
> EXTRAS: I want a hero panel showing biomarker levels over time
Step 3 — Automatic Pipeline
SciFig takes over from here:
- Data Detection — Reads your file, infers column types, detects the scientific domain, builds a
dataProfile - Chart Planning — Recommends chart families (121 types across 13 domains), selects statistical tests, plans panels and palette
- Code Generation — Applies the journal style kernel (Nature / Cell / Science / Lancet / NEJM / JAMA), generates matplotlib code, runs the zero-touch finalizer
- Export — Outputs SVG + PDF figures, source data tables, render QA evidence, stats report, and reproducible code
Step 4 — Receive Output
output/
├── figures/ # SVG + PDF, vector-only, journal-standard dimensions
├── source_data/ # Excel-ready per-panel tables
├── render_qa/ # Layout audit, contract enforcement, overlap report
├── stats_report.md # Methods-section-ready test descriptions
├── code/ # Reproducible generator script + helpers snapshot
└── metadata.json # Provenance, seed, journal profile, palette
| Aspect | Traditional workflow | SciFig |
|---|---|---|
| Tool | Python scripts, seaborn docs, Stack Overflow | One Claude Code skill, natural language trigger |
| Chart choice | Browse examples, guess | Auto-recommend by domain + data shape (121 types) |
| Journal style | Hand-tune rcParams per journal | One token swap (style="nature" → "cell") |
| Statistics | "Just use a t-test" | Auto-select by data shape; refuse unsupported claims |
| Layout | Manual GridSpec per figure | 11 registry-backed layout recipes + narrative arcs |
| Color | Rainbow palette | Wong / Okabe-Ito defaults, colorblind-safe |
| QA | Render → eyeball → hope | Geometric overlap audit + 3-pass finalizer |
| Manual fixes | Drag labels after exporting, then lose reproducibility | Editable SVG is the canonical source |
| CJK | Missing glyph warnings, boxes | Runtime-filtered fallback chain, zero warnings |
| Family | Count | Examples |
|---|---|---|
| Distribution | 14 | violin_strip, raincloud, beeswarm, ridge, ecdf |
| Time series | 11 | line_ci, spaghetti, area_stacked, streamgraph, gantt |
| Matrix / heatmap | 11 | heatmap_cluster, heatmap_triangular, confusion_matrix, dotplot |
| Scatter / embedding | 10 | pca, umap, tsne, scatter_regression, bland_altman |
| Statistical diagnostic | 8 | residual_vs_fitted, cook_distance, qq, leverage_plot |
| Clinical / survival | 12 | km, forest, waterfall, swimmer_plot, nomogram |
| Genomics enrichment | 10 | volcano, ma_plot, manhattan, oncoprint, kegg_bar |
| ML diagnostic | 9 | roc, pr_curve, calibration, training_curve |
| Engineering / spectra | 6 | stress_strain, phase_diagram, nyquist_plot, xrd_pattern |
| Composition / flow | 12 | sankey, alluvial, treemap, sunburst, chord_diagram |
| Ecology / environment | 4 | species_abundance, shannon_diversity, ordination_plot |
| Psychology / social | 4 | likert_divergent, likert_stacked, mediation_path |
| Misc / hybrid | 10 | dumbbell, paired_lines, dose_response, mosaic_plot |
Full registry: runtime/registry.py.
<a id="journal-styles"></a>
| Style | Token | Single column | Double column | Typography |
|---|---|---|---|---|
| Nature | style="nature" |
89 mm | 183 mm | Times New Roman body, 6.5 pt, spine-out ticks |
| Cell | style="cell" |
85 mm | 174 mm | Denser type, story-board layouts |
| Science | style="science" |
90 mm | 190 mm | Minimal axes, narrative-first |
| Lancet | style="lancet" |
84 mm | 174 mm | Clinical conservatism, muted accents |
| NEJM | style="nejm" |
88 mm | 171 mm | Table-grade hygiene, generous whitespace |
| JAMA | style="jama" |
89 mm | 183 mm | JAMA Network typography |
Three corrective passes run automatically before layout audit — generators don't call these explicitly:
| Pass | What it fixes |
|---|---|
_promote_inaxes_text_safety |
Lifts in-axes text to zorder≥20 + white bbox so labels aren't buried under data |
_shrink_heatmap_cell_labels |
Reformats heatmap cell text to fit physical cell width |
| Layout audit | Reports text-vs-line / text-vs-scatter / text-vs-patch overlap > 30% |
Legend contract: exactly one shared bottom-center fig.legend per figure. Outside-right legends and in-axes loc="best" are forbidden by policy and enforced by source-lint.py.
Runtime-filtered fallback chain: DejaVu Sans → Arial → Helvetica → Microsoft YaHei → SimHei → Noto Sans CJK SC → Noto Sans CJK JP → Hiragino Sans. Only installed families survive. Zero findfont warnings.
SciFig/
├── scifig/ # Claude Code / Cursor skill (entire runtime)
│ ├── SKILL.md # Coordinator: gates, phase dispatch
│ ├── reference/ # Progressive disclosure (coordinator details)
│ ├── phases/
│ │ ├── 01-data-detect.md # Phase 1: ingest, domain inference
│ │ ├── 02-recommend-stats.md # Phase 2: chart taxonomy, stats, blueprint
│ │ ├── 03-code-gen-style.md # Phase 3: journal profiles, code generation
│ │ ├── 04-export-report.md # Phase 4: export bundle, metadata
│ │ └── 05-template-distill.md # Optional article-code extraction
│ ├── runtime/
│ │ ├── helpers.py # Finalizer, contracts, layout audit
│ │ ├── template_mining_helpers.py # Kernel, palette, idioms
│ │ ├── registry.py # 121 chart key → generator map
│ │ ├── source-lint.py # Forbidden-pattern lint
│ │ └── generators_*.py # Domain-grouped implementations
│ ├── specs/ # Chart catalog, playbooks, policies
│ ├── knowledge/ # Visual grammar (modules + techniques)
│ ├── resources/ # Palette, layout, zorder JSON registries
│ └── assets/fonts/ # Optional user-supplied fonts
├── .cursor/skills/scifig/ # Cursor project skill (junction → scifig/)
└── docs/gallery/ # README gallery images
Issues and PRs welcome. Committable skill work stays under scifig/.
MIT — see LICENSE. Copyright (c) 2026 Techd.
- 94 reference cases extracted from the local
template/corpus, including open Nature / Cell / Science / Lancet / NEJM / JAMA-style figures. - Color systems credit Bang Wong (Nature Methods 2011) and Masataka Okabe / Kei Ito (JFLY 2008).
- Built as a Claude Code skill.
<a id="中文"></a>
中文版（点击展开）
在 Claude Code 中把实验数据一步转换为投稿级科研图。
SciFig Generate 是一个 Claude Code 技能，可将实验数据（CSV、TSV、Excel 或矩阵文件）转换为投稿级科研图。
把技能放进 Claude Code，对话中提到数据文件即可，由四阶段流水线自动完成数据读取、图表规划、统计检验选择、期刊样式代码生成与导出。无需 pip 安装，所有运行时代码都在 scifig/ 内。
- V0.3.0：纯 skill 仓库 — 移除 pip 包（
src/）、发版脚本（scripts/）、测试和 PyPI/CLI 文档。 - V0.2.1：重构 scifig 技能 — 将 8341 行
generators-distribution.md拆分为 13 个generators_<domain>.py模块；case-index.json瘦身，bulk evidence 移入 gitignoredcase-evidence.json。 - V0.1.2：新增可编辑 SVG 输出，并从 SVG 重新生成一致的 PNG。
- V0.1.1：修复图库样式一致性问题。
- V0.1.0：首次发布。
以下所有图均由 SciFig 使用真实开源数据集自动生成，无任何手工后处理。
SciFig 作为 Claude Code 技能运行，整个流程在 Claude Code 对话中触发——不需要单独的服务器、API 密钥或浏览器标签页。
用户输入触发关键词或 /skill scifig
│
▼
┌───────────────────────────────────────────────────┐
│ SKILL.md — 协调器 │
│ 验证文件路径、读取数据、分派各阶段 │
└─────────────────────┬─────────────────────────────┘
│
┌────────────────┼────────────────┐
▼ ▼ ▼
阶段 1 阶段 2 阶段 3
数据检测 图表规划 代码生成
──────── ──────── ────────
• 验证文件路径 • 按领域和数据 • 应用期刊样式
和编码 形态推荐图表 内核
• 推断数据结构 家族（121 种） • 生成 matplotlib
• 检测科学领域 • 规划 panel 代码
• 构建 dataProfile 布局蓝图 • 运行 finalizer
• 输出： • 输出： • 输出：
dataProfile chartPlan styledCode
│ │ │
└────────────────┼────────────────┘
▼
阶段 4
导出与报告
第 1 步 — 安装
git clone https://github.com/Techd81/SciFig.git
cp -r SciFig/scifig ~/.claude/skills/重启 Claude Code，技能自动被发现。
第 2 步 — 触发
触发关键词：生成图表、画图、科研图、多 panel、generate figure、plot data。或使用：
> /skill scifig
> FILE: /path/to/your_data.csv
> EXTRAS: 我想要一个 hero panel 显示生物标志物随时间的变化
第 3 步 — 获得输出
output/
├── figures/ # SVG + PDF
├── source_data/ # 每个 panel 的 Excel-ready 表格
├── render_qa/ # 布局审计 + 合约执行
├── stats_report.md # 方法学就绪的统计描述
├── code/ # 可复现生成脚本
└── metadata.json # 来源、种子、期刊配置
| 方面 | 传统方式 | SciFig |
|---|---|---|
| 工具 | Python 脚本、seaborn 文档、Stack Overflow | 一个 Claude Code 技能，自然语言触发 |
| 图表选择 | 翻文档、靠猜 | 按领域 + 数据形态自动推荐（121 种） |
| 期刊样式 | 每次投稿手工调 rcParams | 一个 token 切换 |
| 统计 | "随便加一个 t 检验" | 按数据形态自动选检验 |
| 质检 | 渲染 → 肉眼看 → 希望 | 几何重叠审计 + 3 pass finalizer |
SciFig/
├── scifig/ # Claude Code 技能（全部运行时代码）
│ ├── SKILL.md
│ ├── phases/ # 四阶段流水线 + code-gen/
│ ├── specs/ # 图表目录、领域 playbook、期刊样式
│ ├── runtime/ # Python 生成器与 finalizer
│ ├── knowledge/ # 视觉语法知识库（94 案例）
│ └── resources/ # palette、layout、zorder 注册表
└── docs/gallery/ # README 图库图片
欢迎 Issue 与 PR。可提交的 skill 改动请放在 scifig/ 下。
MIT，详见 LICENSE。版权 (c) 2026 Techd。
- 94 个参考案例来自本地
template/语料库。 - 配色系统来自 Bang Wong（Nature Methods 2011）与 Masataka Okabe / Kei Ito（JFLY 2008）。
- 作为 Claude Code 技能构建。