Current public version: 1.0
Paper summary → original prompt template filled without paraphrasing → primary style-faithful PNG → faithful editable PPTX companion → visual-first QA.
中文说明 · Changelog · Contributing
Research Figure Compiler creates Motivation and Pipeline figures for the communication quality expected in leading AI venues such as NeurIPS, ICML, ICLR, AAAI, CVPR, ACL, and KDD. It targets a clear five-second message, source-bounded claims, readable text, precise arrows, balanced density, disciplined alignment, a coherent visual language, and a faithful editable PowerPoint companion. This is a quality target, not automatic venue compliance.
This project can be orchestrated from Codex, Claude Code, and other agentic coding environments that can read the paper, call an image-generation model, create or edit PowerPoint files, render previews, and inspect images. Tool availability differs across environments. When model choice is available, prefer the latest capable GPT model with strong image generation and visual reasoning; other agents may still coordinate the workflow, but cannot produce identical results without equivalent image and presentation tools.
The project is fundamentally a paper-understanding and structure-separation workflow. Its most important asset is the prompt: it reads the article, separates the scientific narrative, and fills the original prompt template without changing its visual-style lock. The complete workflow nevertheless handles much more than prompt writing: paper summarization, structural reasoning, content compression, figure generation, scientific and textual checks, detailed visual review, and an editable PowerPoint companion. This removes many repetitive manual tasks while preserving a clear review path.
Motivation and Pipeline are deliberately separate products:
- Motivation explains why the problem and research need exist.
- Pipeline explains how inputs pass through the proposed method to produce outputs.
Do not merge them into one overloaded figure. If both are requested, generate two separate summaries-to-prompt branches and two separate figures. This separation is a core feature: it reduces prompt complexity, token use, information overload, and scientific-role confusion.
The current Pipeline template still has a relatively structured input → core processing → output bias. Users needing branches, loops, multi-agent interactions, training/inference lanes, detailed algorithm mechanisms, dataset diagrams, comparison figures, or another topology should state those requirements explicitly. The more specific the desired structure, the more specific the request should be.
To save tokens and avoid rework, state these items at the beginning:
- Motivation, Pipeline, or both;
- whether the paper already has a corresponding figure;
- if it does, whether to replace it completely, use it as a visual reference and improve it, or preserve and repair it;
- the desired language, venue, aspect ratio, resolution, required content, forbidden content, and any special topology.
SVG is not a default deliverable. Repeated tests found that adding an SVG generation and inspection path created another visual/text synchronization surface: text and graphics could disagree during review, producing overflow, collisions, or a final image that differed from the approved design. SVG can still be requested explicitly, but the recommended outputs are the primary PNG and a PPTX companion. The PPTX will not perfectly reproduce the PNG's image-model aesthetics; it is intended as a substantially complete, editable template for user revision, while the PNG remains the visual source of truth.
This is the current 1.0 design boundary. More flexible compositions, stronger cross-format fidelity, and additional review capabilities are planned for future updates.
Users should specify:
- Motivation, Pipeline, or both;
- whether the paper already contains the corresponding figure;
- if yes, whether to completely replace it, use it as a visual/layout reference and improve it, or preserve and repair it.
If these choices are missing, the Skill asks one combined question. Complete replacement strictly ignores the old figure and its caption.
flowchart LR
A["Allowed paper content"] --> B["One paper summary"]
B --> C{"Requested figure type"}
C --> D["Complete Motivation master prompt"]
C --> E["Complete Pipeline master prompt"]
D --> F["Primary PNG"]
E --> F
F --> G["Faithful editable PPTX companion"]
G --> H["Visual-first QA of both"]
The Skill does not create evidence ledgers, FigureSpec files, provenance bundles, audit JSON, or multi-round review packages by default.
prompt-templates.md
contains the original user-supplied Research Figure Prompt Template plus
Motivation and Pipeline fill cards:
- Motivation — status quo → observed limitation → bounded research need;
- Pipeline — typed input → 3–7 verb-led stages → typed output.
The fixed formula preserves the original wording and sequence: primary visual/compositional instruction → scientific topic → exact title and components → complete scientific narrative → reference-aligned global layout → exact region proportions/alignment → Negative Prompt. Its reference-lock paragraph and base Negative Prompt may not be paraphrased, shortened, modernized, or reorganized.
The renderer-facing prompt ends after Negative Prompt. PowerPoint,
editability, overflow, and reviewer instructions run afterward and are never
mixed into the image prompt.
The complete filled prompt directly generates the primary PNG, so one image model controls the complete visual system: composition, illustration character, typography hierarchy, icons, borders, arrows, and scientific narrative. After the PNG passes character-by-character and visual inspection, a one-slide editable PPTX is reconstructed faithfully from it. Exact text, numbers, equations, arrows, brackets, braces, and junctions remain editable; any raster illustration layer is disclosed.
For one figure type, only these files remain:
paper-summary.md
motivation-prompt.md or pipeline-prompt.md
motivation.pptx or pipeline.pptx
motivation.png or pipeline.png
SVG is optional only when explicitly requested. Its structural helper remains available for legacy SVG workflows, but SVG inspection is no longer the default QA path.
The primary PNG is the aesthetic reference. The rendered PPTX preview is checked against it for visual fidelity and editability. The PPTX must not redesign the approved PNG into SmartArt or a generic slide diagram.
Post-generation layout and reconstruction safeguards require:
- a 4–6% outer safe area;
- no fixed canvas-occupancy target that forces unnecessary density;
- at least 36 pt slide titles, 26 pt stage titles, 20 pt body labels, and 18 pt connector labels;
- no figure footnotes or ordinary text below 18 pt;
- normally 3–5 stages and one short label per stage;
- zero equations and callouts by default; at most one of each only when indispensable;
- protected title space with no arcs, borders, icons, or connectors;
- content reduction before layout: shorten → remove → move detail to caption → change composition → wrap semantically;
- no clipping, masking, cropping, or z-order trick as an overflow fix.
Both the primary PNG and the rendered PPTX preview are visually inspected. A clean programmatic report does not excuse visible overflow, unexpected wrapping, border contact, clipping, or collision. Semantic overflow also fails: if ordinary labels require zooming, the main route competes with formulas/caveats, or small type is used to preserve a grid, the content or composition must be reduced.
These safeguards may reduce content or repair geometry, but may not redefine the original visual system. Rounded dashed semantic regions, exact reference proportions, hand-lettered headings, simple 2D icons, and the original orange/blue/green accent language remain primary.
Corporate infographics, polished Nature-style vector art, futuristic AI interfaces, gradients, shadows, glossy 3D icons, photorealistic objects, and SmartArt-like redesigns are hard failures.
Every figure declares at most three semantic arrow families, such as primary flow, secondary action, and dashed supervision. Each family fixes:
- color, opacity, width, and dash;
- arrowhead type, width, and length;
- line cap and join;
- semantic meaning.
The shaft and head are one native PowerPoint connector. Separate triangles are not used as arrowheads. Endpoints attach to named shape ports; heads match the shaft, follow its final tangent, and may not show a gap, double outline, overshoot, or shaft-through-head artifact. Connectors are created before nodes so routes remain behind labels and objects.
Before delivery, the Skill:
- inspects every visible word, number, symbol, arrow, and endpoint in the primary PNG;
- runs the presentation overflow test and renders the editable companion;
- views both PNGs at full size and compares their visual systems;
- closely inspects the title, text-heavy regions, all connectors/gutters, and the four outer edge bands;
- fixes and rerenders every unintended overflow, wrap, overlap, clipping, or broken connector.
The six gates are science, structure, text, aesthetics, optics/layout, and arrows/editability. A visible defect fails even when metadata passes. The default stops after the first passing result with at most two targeted repair rounds.
If a paper appears to contain a corresponding figure, the Skill asks before opening it. Users can choose:
- independent replacement;
- reference-led visual/layout improvement;
- preservation and repair.
Complete replacement does not open, OCR, summarize, or pass the old figure or caption to image generation. Reference-led work may preserve visual grammar, but never treats the reference as scientific evidence.
Without a reference, the default remains the original design system: white background, hand-drawn academic infographic style, large black hand-lettered headings, rounded dashed semantic regions, simple 2D scientific icons, restrained orange/blue/green accents, clear hand-drawn arrows, and compact but readable information density.
gh skill install KaiyiHu/ResearchFigureSkill research-figure \
--agent codex --scope userUpdate a tracked installation:
gh skill update research-figure --dir ~/.codex/skillsManual fallback:
git clone https://github.com/KaiyiHu/ResearchFigureSkill.git
cp -R ResearchFigureSkill/skills/research-figure ~/.codex/skills/research-figureReload Codex if an older instruction set remains cached.
Use $research-figure to create both Motivation and Pipeline figures. Fill the
complete master prompt formula, generate the primary PNGs, and provide faithful
editable PPTX companions. The paper has no corresponding figures; generate
independently and stop after visual QA passes.
Use $research-figure to create only a Pipeline figure. The paper has an
existing Pipeline figure; completely replace it and do not inspect its image
or caption.
- Explicitly excluded source regions are never inspected.
- Exact values, equations, axes, arrows, and final labels are deterministic.
- Private or unpublished material is not sent externally without authorization.
- References constrain visual grammar, not scientific evidence.
- The Skill does not replace scientific, statistical, clinical, or legal review.