360 Secure SkillHub
360 安全技能中心

ccf-visual-composer

根据提供的文字、数值、拓扑和参考素材,生成或重排论文机制图、演示或海报用图、视觉表格及数据图;保留数值与结构,检查排版与渲染质量,并按需交付图像、SVG、PDF、PPTX 和绘图脚本。

mikubaka88
热度 23.05
最近更新 2026-10-19 01:40:26
分类:创作依赖:代码执行 / 读写文件 / 第三方 API / 内部调用大模型应用场景:图像生成 / 演示设计 / 研究 / 概念解释 ...

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add mikubaka88/ccfa-skills/ccf-visual-composer
Metadata
name
ccf-visual-composer
description
Render and redesign CCF figures, visual tables, and method/architecture diagrams from supplied content. Use for 绘图美化, 排版, 配色, GPT Image 2 generation, pure SVG, and editable SVG/PDF/PPTX. Preserve values and topology. Experiment evidence design belongs to ccf-experiment-designer; manuscript rewriting and PDF-to-writing exemplars are separate.
metadata
{"ccf_skill_controls":{"handoff_question_mode":"partial","respect_session_denylists":true,"protect_idea_scope_in_writing":true,"private_material_safety":"moderate","shared_controls":"../ccf-common/references/"}}

CCF Visual Composer

Family File Contract

Before file work, reuse canonical paths and one task/artifact working directory; default to project-root ccfa-workfiles/<purpose>/<artifact-id>/. Update in place, preserve inputs, and use UTF-8 with Chinese-text checks. Apply artifact-contracts.md for naming, ownership, and task-created cleanup; handoffs inherit these paths.

Collaboration Contract

Before specialist execution, read and apply ccf-humanization first, then ccf-common. At every handoff, reuse their applicable active rules or refresh missing/changed ones. Both preflights are required even without prose; detailed editing, experiment, and maintenance modes run only when relevant.

Integrate necessary specialist contributions under one owner; reuse valid prerequisites and verify affected results. Follow the cooperation routes without duplicate reports.

Invocation Controls

CCFA Handoff Mode: PARTIAL (Recommended). Follow metadata.ccf_skill_controls.handoff_question_mode, ../ccf-common/references/handoff-modes.md, and ../ccf-common/references/task-modes.md. Reuse shared rules already in context.

For a new scientific method or architecture concept, GPT Image 2 remains the default first-pass renderer unless the user requests pure SVG/code-first output or opts out. Use the verified host capability and its privacy/image instructions. Do not label an unknown backend as GPT Image 2. Existing authorization covers necessary generation and requested reconstruction; offer extra formats only when outside the request.

Core Rule

Create readable visuals from supplied content. Preserve values, units, uncertainty, exact labels, method modules, and typed connections. Quantitative plots use reproducible code. Missing data or topology requires the relevant evidence owner; do not invent it to complete a composition. Preserve scientific terminology and canonical uppercase acronyms; use natural case for ordinary labels.

Classify the destination as paper mechanism figure, presentation/poster, or README/outreach. Paper figures show representations and computation. Choose visual grammar from the method, not a fixed stage-card template.

For new compositions, use references/visual-contract.md: a compact content-fit canvas, shared alignment anchors, scaled gaps, and typography checked at final size. Do not default to a square or force a fixed aspect ratio. Times New Roman is the default; Comic Sans MS serves a requested comic treatment, subject to user/venue typography. For a new visual direction, choose one functional preset from references/adaptive-architecture-style.md. Inspect supplied references, allocate space by explanatory importance, and keep only necessary labels/numbers.

Modes And Selective References

Read only the relevant sections below. A local edit starts from the existing source and applicable QA; it does not reload the complete generation workflow.

Mode Use and reference
visual-contract Non-trivial content/evidence/output decisions: references/visual-contract.md.
figure-design, python-plotting Numerical plots: select from references/python-plot-recipes.md, then import the needed recipe from resources/python/ccfa_plot_recipes.py; read implementation only to debug or adapt it. references/plot-inspiration-map.md is optional for an unresolved chart choice.
architecture-generation New concept: references/architecture-diagram-generation.md; paper-specific grammar only from references/paper-vs-presentation-diagrams.md.
pure-svg-generation Explicit deterministic route: use the supported topology and vector authoring source directly.
editable-reconstruction Semantic reconstruction sections in references/architecture-diagram-generation.md; load references/editable-pptx.md only for PPTX.
reference-layout-blueprint Supplied reference composition: references/reference-layout-blueprint.md; use references/adaptive-architecture-style.md when prompt refinement is needed.
icon-system Native primitives, reusable licensed icons, or necessary custom assets: references/icon-system.md.
table-design, layout-integration Supplied values, panel/float/caption placement: references/figure-table-layout.md.
render-qa Inspect the requested formats using the relevant checks in references/render-qa.md.

Use references/palette-and-accessibility.md only when choosing or changing color semantics. Reuse an established palette and icon family when they remain suitable.

Workflow

  1. Resolve the requested artifact, existing source, final size, destination, formats, and scientific takeaway. For file work, read ../ccf-common/references/artifact-contracts.md once and resolve canonical output and working paths before rendering. Keep unrelated figures in separate stable working directories; honor existing project paths.
  2. Resolve scientific prerequisites before rendering dependent content: source values, units, metric meaning, topology, and intended message. Use ccf-experiment-designer for unresolved result semantics, the method's owner for topology, or ccf-integrity-auditor for source conflicts; integrate their evidence without inventing missing content. Reuse one specification for topology, labels, data locations, layout, style, and provenance. Do not save overlapping contracts, prompt drafts, wireframes, or QA logs. A minor edit reuses unchanged prerequisites and checks only affected dependencies.
  3. Select the smallest rendering route that completes the task. A new architecture concept uses the default image workflow. For an existing SVG, PPTX, or plot script, directly edit that authoring source and export affected requested formats; do not run a new raster concept pass for a label, color, spacing, data, or export change. Raster edits follow the host image-editing workflow and use the existing image as reference.
  4. Preserve a detailed user prompt and add missing constraints once. For a full redesign, include the reference roles, layout geometry, text inventory, typography, and palette needed to make it concrete; do not truncate these to an arbitrary word limit. Reuse approved topology, layout tokens, and assets. Start with one complete candidate. Requested alternatives or Common's applicable figure_variants preference can add a meaningfully different composition using the same approved content; auto requires an unresolved design choice and on applies to new design comparisons, not local edits. Keep each retained candidate in the existing artifact working directory and identify the chosen canonical result. Search or generate assets only to resolve a specific unmet need.
  5. Inspect the draft against the layout and scientific contract. Correct vector/native text and geometry during requested reconstruction; for raster delivery, use a targeted image edit for text, spacing, or illustration defects and preserve accepted regions. If two attempts fail to fix the same defect, diagnose the cause and change the relevant strategy. Never stop merely at an attempt count while a feasible correction remains.
  6. Build requested editable outputs as semantic groups, live text, shapes, and typed connectors. Do not embed a whole raster and claim editability. Keep unavoidable raster assets separate and describe their actual editability. Generate downstream PDF/PPTX from the canonical authoring source without lossy round trips; retain only necessary reusable source and assets.
  7. Check changed outputs for data/topology, whole-figure compactness, shared edges/baselines, final-size text, clipping, contrast, and requested editability. A local edit also checks attached connectors and protected regions. Render affected pages/slides first; broaden only when a changed shared style/layout affects them. Save one current preview under the working build directory; reuse satisfactory inspections.
  8. Finish the requested deliverable and check file placement, current exports, and disposable temporary files. New data/protocol decisions belong to ccf-experiment-designer, manuscript prose to ccf-paper-writer, and claim mismatches to ccf-integrity-auditor; keep unaffected authorized work moving.

Output Contract

Deliver the requested visual and necessary editable source/formats first. A specification-only request needs no render. Return canonical paths, useful provenance, actual QA results, and material editability limits. Keep the full prompt, specification, and inventory in one reusable source record only when needed or requested; do not also repeat them in the final answer. Preserve explicit exact-output requests.