Research figure

Research Figure Compiler

Public workflow version: 1.0.

Use the shortest safe workflow:

allowed paper content
  → resolve whether existing figures may be inspected
  → one useful paper summary
  → complete Motivation and/or Pipeline master prompt
  → primary style-faithful PNG
  → faithful editable PPTX companion
  → visual-first critical check of both

Do not create an evidence ledger, role-analysis file, FigureSpec, provenance bundle, audit JSON, or multi-round review package unless requested.

Resolve SKILL_ROOT as the directory containing this file.

1. Establish the user brief

Target the communication quality expected of leading AI-conference figures: one five-second message, source-bounded claims, reviewer-readable text, precise arrows, breathing room, disciplined alignment, accessible color, and an editable master. This is a quality target, not automatic venue compliance.

Ask for the following missing decisions in one concise question:

Figure type: Motivation, Pipeline, or both?
Existing corresponding figure: yes, no, or unknown?
If yes: completely replace it, use it as a reference and improve it, or
preserve it and repair selected parts?

Do not repeat decisions the user has already supplied.

2. Select the requested output

• motivation: status quo → observed limitation → bounded research need.
• pipeline: typed input → normally 3–5 verb-led stages → typed output.
• Generate both only when the user asks for both.
• Ask before continuing when the type is missing.

Do not classify one type as the winner and suppress the other.

3. Resolve existing figures before inspecting them

Respect exclusions first. Never inspect an excluded figure, caption, page, or supplement.

When the allowed source appears to contain a corresponding figure and the user has not chosen how to treat it, ask:

The paper appears to contain an existing Motivation/Pipeline figure. Should I
(1) regenerate independently and ignore it, (2) use only its visual/layout
language as a reference, or (3) preserve and repair it?

For complete replacement, enforce strict isolation:

• do not open, render, OCR, summarize, or use the old figure or caption;
• do not reuse its topology, palette, wording, icons, or placement;
• if it is already visible in the conversation, explicitly ignore it;
• start image generation without image inputs or recent-image carryover;
• write Reference: none — complete replacement in the prompt.

For reference-led improvement, use the old figure only as permitted visual structure, never as scientific evidence. For repair, preserve verified elements and change only requested or failed parts.

4. Summarize the allowed source once

Create paper-summary.md, normally 500–900 words or equivalent:

1. research problem and importance;
2. current approach and concrete gap;
3. bounded thesis and contributions;
4. method input, 3–7 source-supported stages, handoffs, and output;
5. strongest exact results useful for understanding the paper;
6. limitations and interpretations the figure must not imply;
7. exact terms, numbers, and labels likely to appear;
8. inspected scope, exclusions, and missing material.

Use page or section anchors where practical. Preserve units, qualifiers, uncertainty, and direction. Do not invent missing modules, values, or links.

5. Fill the production prompt

Read references/prompt-templates.md every time. Treat its Original Research Figure Prompt Template as the Skill's immutable core asset. Copy the fixed reference lock, scientific-topic scaffold, reference-aligned layout wording, and base Negative Prompt without paraphrasing, shortening, modernizing, or reorganizing them. Replace only the declared placeholders and delete inapplicable placeholder lines. Fill only the requested Motivation or Pipeline card, copy exact labels from the summary, and list every connector internally as:

source ID → target ID | payload/control label | relation meaning

Save motivation-prompt.md, pipeline-prompt.md, or both. Do not paste the entire paper into the drawing prompt. The renderer-facing prompt must end after the original-style Negative Prompt. Never append PowerPoint, editability, overflow, QA, reviewer, or implementation instructions to it; those are post-generation workflow rules in this file.

Before writing the renderer-facing prompt, derive these fields internally from the summary and then insert them into the master formula:

1. scientific role and the single question the figure answers;
2. five-second message and information priority;
3. source-bounded scientific narrative;
4. necessary visual elements and exact visible labels;
5. forbidden content and claim boundaries;
6. reference visual grammar, when permitted;
7. global composition and region-by-region content;
8. explicit arrow semantics;
9. the unchanged base Negative Prompt plus source-specific forbidden content.

Do not save this derivation as another file. Its purpose is to make the final prompt complete, paper-specific, and visually executable.

Compress content before layout

Every filled prompt must:

• reserve a clean 4–6% outer safe area;
• choose the composition only after reducing the visible-text inventory;
• keep the main title to about 8 English words or 16 Chinese characters;
• use no optional subtitle unless it is essential to the five-second message;
• default to 3–5 pipeline stages; use 6–7 only when each stage needs no more than a short title and one short label;
• keep stage/scene titles to about four words and explanatory labels to about seven words;
• keep connector labels to one to three words;
• show zero equations and callouts by default; allow at most one of each only when essential to the five-second message; move derivations, qualifiers, caveats, repeated explanations, and implementation details to the paper caption instead of a bottom strip;
• use at least 36 pt for the slide title, 26 pt for stage/scene titles, 20 pt for body labels, and 18 pt for connector labels;
• reject all figure footnotes and any non-axis text below 18 pt;
• keep one-line titles on one line;
• shorten wording first, then remove nonessential copy, then wrap at semantic boundaries, then change the composition; never solve overflow by shrinking;
• reserve text boxes before placing icons, connectors, or decoration;
• keep text, equations, arrowheads, and icons fully inside their usable area;
• keep connector labels in open whitespace, never across a border;
• keep unrelated geometry at least 0.5 body-line height from text;
• use brackets/braces only for declared grouping targets and junctions only for source-supported convergence.

Preserve the original visual system

Do not let overflow safeguards redefine the style. The original prompt's visual language remains primary:

• white background;
• hand-drawn academic infographic style;
• large black hand-lettered title with slightly irregular strokes;
• rounded dashed semantic regions in restrained orange, blue, and green;
• simple consistent 2D scientific icons;
• compact but readable information density;
• clear hand-drawn arrows;
• no gradients, shadows, glossy 3D, corporate UI, or Nature-style polished vector redesign.

Use the permitted reference's topology, proportions, region borders, gaps, icon scale, arrow rhythm, density, and typography hierarchy. Do not replace rounded dashed regions with an open canvas merely because a generic layout rule prefers it. Without a reference, choose two or three broad regions in this same original visual system and state their exact proportions and alignments in the prompt.

Do not add a bottom formula ribbon, disclaimer bar, or repeated summary callouts unless the science requires it. Do not target a fixed canvas-occupancy percentage, but do specify the region proportions required by the original prompt formula.

Treat the main title as a protected lane. Keep panels, arcs, connectors, borders, icons, and texture outside it. Start body content below the title lane plus at least 0.5 body-line height. Do not hide a collision with a mask, crop, or z-order trick.

Arrow system contract

Declare an ARROW_STYLE_TOKENS table before assembly. Use no more than three semantic families:

FLOW_PRIMARY   | color | 2–3 px | solid  | arrow/stealth med | round | round
FLOW_SECONDARY | color | 2–3 px | solid  | arrow/stealth med | round | round
SUPERVISION    | color | 2–3 px | dashed | arrow/stealth med | round | round

Omit unused families. A token fixes stroke color, width, dash, opacity, arrowhead type/width/length, line cap, line join, and semantic role. The same meaning must use the same token everywhere.

For every arrow:

• use one native connected object for the shaft and head;
• never assemble a separate triangle and line;
• make head fill/stroke match the shaft;
• attach endpoints to named shape ports, not approximate floating coordinates;
• keep the head tangent to the final shaft segment with no gap, double outline, overshoot, or shaft visible through the head;
• create connectors before nodes so lines remain behind nodes and labels;
• route around all text boxes and keep labels beside the route in reserved whitespace;
• use dash/color to distinguish semantics, not arbitrary head shapes;
• render and visually inspect every bend, endpoint, crossing, and merge.

6. Generate the primary PNG and editable companion

Use the Presentations Skill for every PPTX workflow and follow its local instructions. Use @oai/artifact-tool from a JavaScript ES module; do not use python-pptx.

Default deliverables are:

motivation.pptx + motivation.png
pipeline.pptx + pipeline.png

Use one slide per figure. Keep all final titles, labels, numbers, equations, panels, arrows, leaders, brackets, braces, and junctions as native editable PowerPoint objects. A stylistic illustration layer may remain raster only when disclosed; never call a flattened bitmap fully editable.

Use aesthetic-first image generation by default, including when no reference exists. Send the complete filled original prompt—not a shortened summary and not a rewritten meta-prompt—to the built-in image-generation tool. Do not append the downstream PowerPoint or QA instructions. Generate the primary PNG as a complete scientific figure with the declared title, labels, arrows, icons, panel topology, and visual language. This restores the original prompt-driven workflow in which one coherent image model controls the whole design system. In reference-led improvement or repair mode, include only the user-permitted reference image with the complete prompt so its observable visual grammar can control the result. In complete-replacement mode, include no reference image and no recent-image carryover.

Inspect the primary PNG before PowerPoint reconstruction. Reject or revise it if it looks generic, childish, corporate, diagram-template-like, visually unbalanced, misspelled, structurally wrong, blurred, or overcrowded. Check every visible word, number, symbol, equation, arrow, and endpoint character by character. Never accept pseudo-text merely because the composition is good.

After the primary PNG passes, create the PPTX as a faithful editable companion using the approved PNG as the visual reference. Reproduce its composition, scale relationships, border language, icon placement, arrow rhythm, palette, and typography hierarchy rather than redesigning it. Keep exact text, numbers, equations, arrows, brackets, and junctions editable. A raster illustration layer may remain and must be disclosed.

Use native-first assembly only when the user explicitly prioritizes fully editable deterministic geometry, or when exact plots/equations dominate. Native-first must still be designed from an approved visual concept and must not fall back to default ovals, cards, SmartArt, or clip art.

When no reference exists, use a publication-grade hand-drawn academic infographic by default: white paper, refined slightly organic black linework, large black hand-lettered headings, clean readable body labels, concrete scientific 2D icons, restrained orange/blue/green accents, rounded dashed semantic regions, clear hand-drawn arrows, and compact but readable information density. Do not substitute Comic Sans, default PowerPoint shapes, corporate SmartArt, polished Nature-style vector art, or a futuristic AI interface.

Create connectors before entity nodes with slide.shapes.connect(...). Prefer side anchors or explicit connection sites. Use the token's built-in head, line, cap, and join properties so the arrowhead and shaft remain one consistent PowerPoint connector.

SVG is not a default deliverable and is not the default QA surface. Create it only when the user explicitly requests SVG.

7. Run visual-first QA

The approved primary PNG is the aesthetic reference. The rendered PPTX preview is the editability/fidelity reference. Passing metadata or code checks never overrides a visible defect in either.

For each PPTX:

1. run the Presentations Skill's slides_test.py;
2. render every final slide with render_slides.py;
3. inspect both the primary PNG and PPTX-rendered preview at original resolution with an image-viewing tool;
4. compare their global topology, scale, typography hierarchy, icon language, border treatment, palette behavior, and arrow rhythm;
5. inspect the title, each text-heavy panel, every gutter/connector route, and all four outer edge bands at a close view;
6. fix every unintended overflow, wrap, overlap, broken connector, crop, or unacceptable style drift;
7. rerender and reinspect after any fix.

Do not approve a figure unless all six gates pass:

1. Science — no invented or stronger-than-source claim, number, or module.
2. Structure — entities, order, endpoints, direction, branches, brackets, and junctions are correct.
3. Text — exact glyphs and numbers; no pseudo-text, unexpected wrap, clipping, border contact, or collision.
4. Aesthetics — looks like a publication figure, not PowerPoint SmartArt, a classroom worksheet, generic clip art, a corporate process diagram, or a collection of oversized pastel shapes. Typography, illustration, palette, and stroke character form one deliberate visual system.
5. Optics/layout — no blur, melted shape, overlap, off-canvas content, edge crowding, tiny auxiliary copy, information overload, repeated narrow cards, dense bottom strips, inconsistent gap, or misalignment.
6. Arrows/editability — consistent token use; no floating endpoint, shaft/head gap, mismatched head, broken bend, connector through text, or flattened final text/structure.

Programmatic warnings are a lower bound, not a waiver. A visible defect fails even when slides_test.py is clean. Conversely, inspect a reported overlap before deciding whether it is intentional.

Treat semantic overflow as a failure even when every text box is technically inside the slide. Fail when a reader must zoom to read ordinary labels, when the main route competes with caveats or formulas, or when the composition uses small type to preserve a preselected grid. Remove content or change the composition rather than making the text fit.

If a gate fails, make one targeted repair and inspect again. Stop at the first passing result. Use at most two repair rounds unless the user requests more.

For explicitly requested SVG only, scripts/quick_qa.py remains an optional legacy structural check. It does not replace PPT/PNG visual inspection.

8. Keep delivery small

For one requested type, keep only:

paper-summary.md
motivation-prompt.md or pipeline-prompt.md
motivation.pptx or pipeline.pptx
motivation.png or pipeline.png

Reuse one summary for both figures. Report QA in the final response instead of creating a QA file. Remove temporary source modules, renders, crops, layout JSON, and montages after approval.

Boundaries

• Do not expose private or unpublished material to an external provider without authorization.
• Keep exact values, equations, axes, and high-risk labels deterministic.
• Do not claim venue compliance without checking current official rules.
• Do not closely copy a reference figure or imitate a living artist.
• Do not replace expert scientific, statistical, clinical, or legal review.