Academic Figure Designer and Spec Engine

Design evidence-grounded academic figures, make semantic color & surface decisions, and compile clean academic-figure/FigureSpec@1 specifications and normalized structured rendering briefs. This skill serves as the single authoritative designer and prompt compiler for all figure styles.

Load only as needed:

• spec contract → json-schema.md and figure-spec.schema.json
• prompt compilation → references/json-to-prompt.md
• visual brief guidance → references/image-prompt-guide.md
• composition scaffolds → references/prompt-templates.md
• visual anchors & SVMC → references/architecture-icons.md
• palette/style fallback → references/palettes.md and optional references/styles/
• missing evidence → references/missing-info-policy.md

Supported Style Profiles

Select one canonical style_profile (or compose with layer overlays from docs/styles/):

1. classic-technical (经典学术框线风):
   • Clean white background, thin 1.5pt crisp outlines, restrained subtle tints (Okabe-Ito, Nature Blue, Cool Gray).
   • High contrast, orthogonal alignment, strict box/arrow engineering topology.
   • High-density tabular parameters and formal sans-serif typography (Helvetica/Inter).
2. pastel-airy-ui (现代柔彩空气风):
   • White canvas with floating white cards and faint borders.
   • Generous negative space, floating pills/tokens (P1 Warm, P2 Cool, P3 Earthy), and lightweight curves.
   • Interface-like feel without multi-level nested boxes.
3. illustrated-modular (编辑手绘模块风 / 有色语义分区图示风):
   • White canvas with content-driven soft-tinted semantic zones (I1 paired tokens: {soft_fill, dark_outline, title_text, icon_accent}).
   • Strong same-hue 1.5–2.5px dark outlines and no drop shadows.
   • Asymmetric hero region (~35–55% visual focus) with supporting modules arranged by real semantics.
   • Content-grounded editorial line art tied to declared semantics.
4. reference-led (参考图驱动自由风格):
   • Extracts observable visual grammar (composition, marks, stroke, typography, illustration level, paired color fills/outlines) directly from a user-supplied reference image without copying proprietary content or topology.

Direct Input Normalization Protocol (自然语言直通归一化)

When the user provides direct natural-language architecture descriptions (bypassing upstream analyzers), deterministically normalize the input into FigureSpec v1:

1. Source Mapping:
   • sources: [{"kind": "user_instruction", "uri_or_path": "conversation", "evidence": "<exact user request text>"}]
   • plan_revision: "r1"
2. Topology Synthesis:
   • Extract stages/modules as snake_case IDs (input_data, encoder, fusion_block, loss_head, output).
   • Derive directed connections from temporal or dataflow verbs (e.g. "送入", "经过", "transforms to").
   • Default must_not_claim: [], forbidden_connections: [], and standard negative_constraints.

Semantic Color & Accessibility Decision (Domain-Adaptive Presets)

The default rules below serve as recommended presets for ML, AI4Science, and Systems papers. Domain-specific figures (e.g. monochrome print, inverted microscopy, multi-channel bio, astrophysics) may customize palette and background while preserving readability:

1. Recommended ML / Systems Role Binding (Preset):
   • Multi-stage pipeline: Stage 1 (Green input) -> Stage 2 (Blue encoder) -> Stage 3 (Peach transformation) -> Stage 4 (Purple loss) -> Stage 5 (Gold output).
   • Deep Learning: Features/Data (Green) -> Backbone (Blue) -> Fusion/Attention (Peach) -> Loss/Constraint (Purple) -> Head/Task (Gold).
   • Interactive / Agentic: Policy/Reasoning (Blue), Context/Observation (Green), Tool/Harness (Peach), Advisory/Feedback (Purple), Memory/Storage (Cyan), Output (Gold), Guardrail/Stop (Coral).
2. Domain-Specific Extensions:
   • Monochrome Print: High-contrast grayscale tints (#FFFFFF fill, #24323D stroke, dashed vs solid line dual-encoding).
   • Dark/Microscopy / Astrophysics: Dark canvas allowed when reference or domain data requires inverted emission contrast.
3. Accessibility & Grayscale Invariants:
   • Ensure title and label texts meet WCAG AA contrast against their card backgrounds.
   • Dual-encode critical paths with shape, border style (solid vs dashed), or iconography so meaning survives grayscale printing and color vision deficiencies.

Strict Prompt Formatting Standard (Prose Normalization)

All generated image prompts MUST be compiled as normalized structured natural language (Compact Prose):

[!IMPORTANT] Zero Markdown Syntax in Image Prompts: Never use Markdown formatting symbols (such as # headers, **bold**, *italic*, markdown bullet lists - item, backticks, or ASCII markdown tables | --- |) inside the prompt string sent to diffusion/image models. Image models frequently hallucinate and render Markdown syntax tokens as literal text on the canvas. Use clean, comma-and-sentence structured prose grouped by numbered container blocks or semantic zones.

Canonical Prompt Structure

1. Lead & Purpose: High-level figure goal, aspect ratio, canvas background (pure white #FFFFFF), and style profile. Default to an external paper caption with no canvas title. If the user, FigureSpec, or supplied reference explicitly requires a title, lock exactly one short non-banner title and reserve whitespace for it.
2. Layout & Hero Focus: Numbered container panels and proportions (e.g. 3-column sandwich layout, central hero region).
3. Semantic Container Blocks: For each container, describe inner title pill, sub-cards, data flow, and scientific visual metaphors (SVMC).
4. Topology & Connections: Explicit source -> destination connections, line styles (solid for forward, dashed for feedback/advisory), and edge labels.
5. Visual Constraints: Negative defect constraints (No unintended or duplicated title banner, no floating text, no gradients, no 3D chrome, no photorealism, no shadows).

Text Budget

Visible text must remain structural and publication-legible:

Element	Guideline
Region or module title	usually no more than 5 words
Short label	usually no more than 3 words
Arrow label	usually no more than 3 words
Core formula	at most one short sourced line
Parameters, evidence, caveats	caption only

Drop secondary labels before reducing them below readable final-paper size.

Output Contracts

This skill emits one of two deliverables based on the user's intent:

A. Full Figure Design (Default)

Emits FigureSpec v1 JSON and normalized compact prose prompt. Proceed to validate and render.

B. Palette Decision (Color / Consultation Only)

When the user asks only for color palette advice, style recommendation, or accessibility review (学术配图配色, 论文配色方案, 色盲友好配色): Emit a structured Palette Decision and stop:

• Canonical style_profile and decision branch (user, reference, scene, or default);
• Recommended palette / paired token set (and one alternate);
• Canvas, body text, outline, and neutral divider colors;
• Semantic-zone bindings mapped to the paper's actual roles;
• Accessibility & grayscale dual-encoding notes;
• Copy-ready token handoff.

Workflow

1. Identify Task Scope (Design vs Palette-Only)

• If the request is palette/color/style consultation only, perform Step 3 (Bind Style & Color Tokens) and emit the Palette Decision, then STOP.
• If the request is full figure creation or prompt compilation, proceed through all steps.

2. Close the Semantic Graph & Geometry

Copy component IDs, labels, groups, and typed connections from upstream analysis or direct user input. Carry must_not_claim, forbidden_connections, and authority boundaries into the spec.

3. Choose Composition & Visual Metaphor (SVMC & Spatial Blueprints)

• Spatial Container Allocation: Express macro-containers with explicit canvas height/width percentages (e.g., 2-tier stacked containers with 30-35% top vs 65-70% bottom, or left-hero 40-50% vs right-stack 50-60%) to prevent empty canvas dead zones.
• Nested Card Scaffolds: Use outer macro-containers with subtle dashed borders and nest solid white sub-cards inside.
• Micro-Visual Trinity Injection: Never output empty blank boxes. For each key node, compile the Micro-Visual Trinity from references/json-to-prompt.md and references/architecture-icons.md:
  1. Header / Icon Badge (domain icon, e.g. 💡, 🔬, 🔍, 📊);
  2. Concrete Scientific Schematic (3D GP elevation mesh, 1D multi-peak acquisition curve, 3D tensor block, heatmap, or state DAG);
  3. Micro Mathematical/Data Card (equation card, uncertainty gauge, dialogue bubble, or mini table).

4. Bind Style & Color Tokens Semantically

Bind paired tokens to semantic regions: background, soft fill, dark outline/title, optional icon accent, and exception color from docs/palettes.md. Follow dual-fidelity and multi-role paired color rules (e.g. Coral Red for real/high-fidelity vs Slate Blue for surrogate/low-fidelity).

5. Emit FigureSpec v1

Conform to figure-spec.schema.json. Required features include:

• style_profile: classic-technical, pastel-airy-ui, illustrated-modular, or reference-led.
• unique component IDs, closed visible-text list, sources, typed connections with valid endpoints.
• prompt_review: requested|confirmed|waived and conditional prompt_reviewed_sha256.
• declared absolute workspace_root and output_path.

6. Validate FigureSpec v1

Immediately before every render or edit, run:

python3 academic-figure-designer/scripts/validate_figure_spec.py \
  --strict-v1 --render-ready \
  --workspace-root <trusted-actual-root> \
  <spec.json>

7. Hand off for Rendering & Repair

When called from academic-figure-workflow, pass the validated rendering package forward. Use the current session's native image_gen.imagegen interface. If prompt review is waived, keep prompt internal as a tool parameter and deliver the rendered image directly.

Stop Condition

• Palette-only requests: Stop after emitting the complete Palette Decision. Do not demand topology or attempt image rendering.
• Figure design requests: Stop after emitting and validating FigureSpec v1 (or delivering the rendered image in workflow mode).