Academic Figure Prompt

Academic Figure Prompt

Default deliverable: a JSON figure spec (exact_* text locks + layout blocks + rendering rules). Text prompts only for simple charts or explicit user request.

Schema and examples: → json-schema.md
Palettes: → references/palettes.md
Image prompt writing: → references/image-prompt-guide.md
Icon vocabulary: → references/architecture-icons.md
Prompt templates: → references/prompt-templates.md
JSON→prompt conversion: → references/json-to-prompt.md
Missing info: → references/missing-info-policy.md

Text Budget (leading rule)

On-figure text is short labels and structure. Formulas, params, and long prose go to Figure Caption.

element	limit
module title	≤ 5 words
subcomponent	≤ 3 words
pipeline step	≤ 2 words primary + ≤ 2 secondary
formula on figure	≤ 1 line core only
arrow label	≤ 3 words

Label hierarchy: Primary (must read at a glance) → Secondary (drop first under space pressure) → Caption (never on figure).

Input Contract

• Prefer: figure type, paper/section content, modules, labels, formulas, dims, Palette Decision, reference image

• aspect_ratio (optional, from Figure Plan)

• Minimum: figure type + subject/method overview

• Missing: skeleton spec with placeholders; mark 推断 / 待确认

• Chinese figure name + type

• JSON spec (structured intermediate)

• Image prompt (200-400 word English visual brief for the image model)

• palette name + hex used

• caption reserve list

• completeness block (see missing info policy)

Steps

Step 1: Ground content

Read available paper/section material. Extract modules, dataflow, symbols, dims.

Done when: every claimed module has a source span or is marked placeholder.

Step 2: Reference image (if any)

Extract palette, layout flow, box style, annotation density, special links.

Done when: reference constraints are listed or “no reference” is explicit.

Step 3: Palette

Do not maintain a private palette table. This skill is classic family only (pastel → other skill).

1. User-specified palette / hex → use it
2. Else existing Palette Decision from color-expert → use it
3. Else run references/palettes.md Scene → palette decision (hard constraints → type → venue → domain); if still empty, safe default (≥4 modules → Nature Blue; else Okabe-Ito) and say so
4. Load hex from references/palettes.md
5. If user signals airy/pastel, stop and route to academic-figure-prompt-pastel instead of forcing classic borders

Done when: palette name + hex are fixed, family is classic, and the decision branch is stated.

Step 4: Emit JSON spec

Load json-schema.md. Build layout_and_content_blocks with exact_* locks for every visible word. Include physical_spec_and_typography (canvas 89mm/183mm, font hierarchy 10pt/8pt/6pt, stroke hierarchy 1.5pt/1.0pt). White fill + colored borders. Attach rendering rules and caption_note list.

Done when checklist passes:

• every on-figure string is in an exact_* field
• aspect_ratio copied from Figure Plan when present
• physical_spec_and_typography block present (89mm/183mm width, font 10pt/8pt/6pt, stroke 1.5pt/1.0pt)
• Text Budget respected
• white fill / colored borders only
• ≤ 3 chromatics from chosen palette
• caption reserve lists off-figure content
• no empty module shells
• every major block has an icon or visual anchor (see references/architecture-icons.md)
• weight status (frozen vs trainable) uses non-emoji pattern (dashed/solid borders, hatching, or pills)
• explicit negative instructions included: NO emojis, NO lock/fire/lightning icons, NO 3D rendering

Step 5: Write image prompt

Read references/image-prompt-guide.md and references/json-to-prompt.md. Convert the JSON spec into a 200-400 word English image prompt following the 8-slot structure:

1. Image type (lead with this)
2. Core subject (one sentence)
3. Composition/layout (spatial arrangement, flow, grouping)
4. Supporting modules (icons from architecture-icons.md, dimension labels, formulas, token pills, legends)
5. Visual tone (concrete descriptors, not vague words)
6. Material/texture (border width, fills, corner radius)
7. Typography & Physical Specs (font hierarchy: title 10-12pt bold, label 8-9pt, tensor 6-7pt; column width: 89mm single / 183mm double; stroke width: 1.5pt borders, 1.0pt dividers)
8. Aspect ratio (last)

Use references/prompt-templates.md for the template matching the figure type. Every major block must have a visual anchor (icon, thumbnail, or geometric marker). Parameters and long formulas go in caption_note, not on the figure.

Done when: prompt is 200-400 words, all 8 slots present, every JSON block translated to spatial prose (not listed mechanically), and supporting modules included.

Step 6: Fallback text prompt (rare)

Only if ≤ 3 modules without branches, pure data chart, or user demands prose prompt. Use four-layer skeleton in json-schema.md (Global Context → Section/Column Encapsulation → Annotations & Links → Style Specifications with hex & negative constraints).

Stop

Stop when the Figure Spec Package (JSON spec + image prompt) for the requested figure(s) is delivered, or when figure type and subject are both missing (ask for those two only).