An end-to-end multi-stage pipeline that extracts figures from scientific PDFs, analyzes them with Kimi K2.6 multimodal vision AI, and reverse-engineers their generation workflows.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install 9cf703b
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install 9cf703b using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Scientific Figure Analysis?
The Scientific Figure Analysis tool is a robust multi-stage pipeline designed for AI agents operating within Openclaw Skills to parse complex, multi-panel figures from scientific literature (such as Nature, Cell, and Science papers). By combining programmatic PDF extraction with state-of-the-art vision models, it bridges the gap between raw pixel data and structured scientific knowledge.
This skill empowers researchers and developers to automatically dissect charts, microphotographs, plots, and pathway diagrams. It goes beyond simple OCR by performing layout analysis, visual grouping, and multi-modal reasoning to output either highly structured JSON data for downstream software integration or clear Markdown tables optimized for direct human review.
Scientific Figure Analysis Use Cases
- Dissecting and analyzing complex, multi-panel figures in scientific and academic PDFs.
- Extracting structured scientific knowledge and numeric data directly from visual charts and plots.
- Reverse engineering the exact software, libraries, and pipelines (such as ggplot2, GraphPad Prism, or Seurat) used to generate figures.
- Automating high-throughput literature mining and database curation from preprints and published papers.
How Scientific Figure Analysis Works
- Extraction (Stage 1): Programmatically extracts all embedded images and text blocks from the PDF using PyMuPDF and pdfplumber, preserving original resolution, layout boundaries, and captions.
- Visual Analysis (Stage 2): Submits extracted images to Kimi K2.6 Vision AI via the Moonshot API to detect and interpret data points, labels, and trends within the figures.
- Scientific Understanding (Stage 3): Classifies the figure into one of 18 specific scientific categories (such as UMAP, Volcano plots, Western Blot) and performs multi-modal reasoning combining image contents with surrounding body text.
- Reverse Engineering (Stage 4): Infers the originating visualization software and reconstructs the data processing workflow (such as normalization or clustering steps).
- Structured Delivery: Generates the final output in machine-readable JSON or executive-ready Markdown tables depending on the workflow requirements.
Scientific Figure Analysis Setup
To configure this skill within your Openclaw Skills environment, ensure you have the required Python libraries and external OCR dependencies installed.
First, install the core Python parsing and computer vision packages:
pip install pymupdf pdfplumber opencv-python layoutparser
Next, install Tesseract OCR for physical text extraction from complex rasterized images (if layout boundaries cannot be extracted programmatically):
sudo apt install tesseract-ocr
Finally, set up your Moonshot API key as an environment variable to enable Kimi K2.6 multimodal vision capabilities:
export MOONSHOT_API_KEY="your-api-key-here"
Scientific Figure Analysis Data Schema & Taxonomy
This Openclaw Skills integration structures its processed visual and textual metadata using two distinct targets: a machine-readable JSON schema and human-readable Markdown templates.
JSON Schema (Format A)
Designed for downstream automated agent pipelines:
| Field Name | Type | Description |
|---|---|---|
figure_id |
String | The identifier of the figure (e.g., "Figure 3") |
caption |
String | Full raw caption text parsed from the PDF context |
overall_summary |
String | Core high-level scientific finding represented |
subfigures |
Array | Detailed objects representing panels a, b, c, etc., including detected labels, chart types, and numeric values |
Human-Readable Markdown (Format B)
Designed for direct executive review, featuring:
- Structured Summary Table: Outlining panel counts, experiment modalities, and key insights.
- Software Inference & Confidence Scores: Categorized as high, moderate, or speculative.
- Reverse Engineering Flowchart: Linear step-by-step markdown block showing the estimated generation pipeline.
Scientific Figure Analysis Advanced Features
- Intelligent Multi-Panel Splitting: Automatically detects and segments multi-panel figures (a, b, c, d) while preserving axis labels, shared legends, and local annotations.
- Fallback Text-Only Pipeline: Seamlessly switches to a highly optimized regex and textual semantic fallback pipeline when vision APIs are unavailable or rate-limited.
- Software Signature Inference: Analyzes typographic details, default color palettes (such as ggplot2 grey grids or Prism Helvetica axes), and formatting cues to accurately identify generating software platforms.
- Conservative Confidence Mapping: Applies strict validation rules to flag uncertain visual details, categorizing observations into high, moderate, or speculative confidence tiers to eliminate AI hallucinations.
- Seamless Agent Native Workflows: Built to integrate natively with Openclaw Skills, enabling developers to connect raw PDF parsers with downstream vision pipelines effortlessly.
SKILL.md
Loading
Related Openclaw Skills
CodeGraph Assistant
An automated wrapper for npm codegraph that simplifies codebase indexing, structural search, impact analysis, and context injection for AI programming.
wutrix
An intelligent film pre-production toolkit designed to seamlessly capture creative ideas, manage scripts, organize character details, and research worldviews directly from your AI workspace.
Artifact Gallery
Converts local files generated by headless AI agents into secure, mobile-accessible HTTPS viewer links.
OpenRouter Rankings Screenshot & Lark/Feishu Digest
An automated workflow tool for Openclaw Skills that scrapes OpenRouter rankings, generates summarized digests, and publishes structured reports with screenshots directly to Lark/Feishu.
WeChat Mini-Program Vibe Coding Starter
An AI agent skill designed to help non-engineers and product managers build, iterate, and launch WeChat Mini-Program MVPs from scratch using Vibe Coding methodologies.
PDF to Structured JSON
A zero-configuration, intelligent document parsing skill powered by Laiye ADP that converts PDFs, images, and Office documents into structured JSON layouts.
Featured*
