Communitygithub.com

academic-figure-design

Design and revise academic flowcharts and schematics, including algorithm/control block diagrams and scene/platform illustrations; reconstruct approved concepts as editable PPT/vector sources and verify manuscript placement. Use for 科研绘图、论文插图、流程图、框架图、示意图、状态机图、实验平台图、结果图排版、AI生图与PPT复现. Not for ordinary slide styling or inventing experimental evidence.

academic-figure-design 是什么?

academic-figure-design is a Claude Code agent skill that design and revise academic flowcharts and schematics, including algorithm/control block diagrams and scene/platform illustrations; reconstruct approved concepts as editable PPT/vector sources and verify manuscript placement. Use for 科研绘图、论文插图、流程图、框架图、示意图、状态机图、实验平台图、结果图排版、AI生图与PPT复现. Not for ordinary slide styling or inventing experimental evidence.

兼容平台~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/dangerous090566/academic-figure-design-skill/tree/main

在你喜欢的 AI 中提问

打开一个已预加载此 Agent Skill 的新对话。

文档

Academic Figure Design

Produce figures that explain a scientific mechanism, not decorated paragraphs. Preserve technical meaning, the user's latest edits, and the distinction between a concept preview, an editable source, and a manuscript-ready figure.

Establish the Task Boundary

Identify the requested stage: concept design, targeted revision, editable reconstruction, or manuscript insertion. Do only the authorized stages. In particular, discussing a new figure does not authorize replacing it in the paper.

Inspect the supplied figure, authoritative manuscript definitions, and accessible source assets before designing. Treat a style example as a visual reference, not as a source of technical modules or claims. Resolve differences using the user's declared source authority; report genuine contradictions rather than silently changing the method.

For an existing project, identify authority separately for scientific content, visual edits, and manuscript placement. A user-edited PPT normally supersedes its older PNG or generation script for appearance, but does not silently override experimental data or method definitions. Preserve its edits and use the existing presentation/document when requested. A timestamp helps detect changes; it is not proof that two files describe the same revision.

Choose between Two Visual Forms

Use references/figure-types.md to choose between flowcharts (流程图) and schematics (示意图). Flowcharts use functional blocks and directed connections to explain computation, decisions, or control/signal flow. Schematics use physical or conceptual objects, geometry, scenes, photographs, and local annotations to explain relationships, mechanisms, or events. Algorithm frameworks, state machines, and simulation block diagrams usually take the first route; explanatory system views, traffic sequences, and photo-led platform views can take the second.

Classify by the dominant visual expression, not by the figure's topic, title, number, or the mere presence of arrows. A schematic can contain a small control-chain inset; a flowchart can contain vehicle miniatures. These do not change its primary form. Modeling, platforms, scenarios, and results are application contexts, not additional peer categories. Data-derived plots still require the quantitative production rules below, irrespective of the surrounding layout.

Write one sentence stating what this figure must let the reader understand. A platform figure explains execution and feedback, not the whole proposed algorithm again. A state-to-mode mapping need not repeat the same traffic scenes already used to define its predicates. Keep each fact at its most informative location.

Semantic Contract Before Layout

Make a compact working inventory of:

  • Modules and their actual roles.
  • Inputs, outputs, and arrow source/target for every signal.
  • Parallel branches, decision conditions, and mode-dependent settings.
  • Mathematical symbols, hard/soft conditions, and what each visual encoding means.
  • What is schematic, measured, simulated, or photographed.

Do not infer a signal connection merely because two modules sit next to each other. Separate a model used to construct constraints from the plant executing commands. Separate a penalty weight from the resulting action. A state-machine motif is not a verified transition graph.

Choose the Production Route

  • Concept or AI revision: Read references/design-and-ai.md. Use the available image-generation skill/tool when the user requests AI imagery. Review the resulting pixels and connections, not only the prompt.
  • Editable PPT/vector reconstruction: Read references/editable-and-insertion.md. Follow the available presentation skill for file operations. Whole-image insertion is not editable reconstruction.
  • Manuscript insertion: Read the insertion section of the same reference and use the relevant document/PDF skill. Change surrounding prose only where the new figure requires a precise explanation.
  • Quantitative plots: Use original data and deterministic plotting tools. AI may help conceptual composition, but must not generate or alter curves, samples, metrics, or experimental evidence.
  • Platform figures: Use genuine supplied photos. Clearly distinguish physical devices from virtual counterparts. Missing equipment details or photos require a specific request, not invented assets.

Visual Defaults

Prefer informative graphical structure over both text-heavy flowcharts and overly sparse office diagrams. Use vehicles, geometry, state nodes, feasible sets, or other domain-relevant graphics to explain the mechanism. Keep short labels and only essential equations; put full definitions in captions or prose. For a local revision, preserve approved regions and the existing style unless a technical conflict requires attention.

Align the composition to an invisible outer rectangle while retaining clear gutters. Reserve enclosing frames and dashed groups for meaningful functional boundaries, not every item. Scenario illustrations, trajectory panels, and platform photographs usually need whitespace and plain subfigure labels instead. Keep typography, vehicle symbols, and semantic colors consistent across types without forcing identical containers. Within a scenario overview, a model-comparison panel can use the same road-scene grammar, with comparison details in its caption. Use white backgrounds, restrained functional colors, readable dark text, and consistent lines and arrowheads. Color must not be the only carrier of meaning.

Design at the intended publication size. For a double-column figure, approximately 180 mm width and 8-10 pt final labels are useful starting points, subject to the actual template. Split a dense figure rather than shrinking everything. Keep figure number and full caption outside the artwork.

Use a small typography system: one English text family with fixed heading/body roles, a compatible math family with real scripts, and a CJK family only when needed. Do not force mathematical variables into ordinary text styling. Compare the whole figure set, not only the slide being edited.

Verify Before Delivery

Use references/verification.md for the final semantic, visual, editability, and insertion checks. Inspect both the full figure at publication size and detailed crops. Recheck unaffected connections after an AI edit because a local request can change distant content.

For recurring failures or skill maintenance, read references/iteration-lessons.md: it turns revision feedback into defect-specific acceptance tests rather than accumulating vague style preferences.

Deliver the requested preview and actual source/export files with concise change notes. Describe limitations honestly: raster previews, schematic-only transitions, unavailable photos, unverified hardware, or untested exports. Do not claim a figure is editable, technically verified, or inserted unless those steps were performed. Keep earlier approved versions recoverable without proliferating new working documents.

相关技能