ccf-visual-composer
CCF Visual Composer
Invocation Controls
CCFA Handoff Mode: PARTIAL (Recommended). Follow metadata.ccf_skill_controls.handoff_question_mode and ../ccf-common/references/handoff-modes.md.
Treat image generation and editable reconstruction as two separate stages. For method, architecture, system, pipeline, framework, and illustrative scientific diagrams, use GPT Image 2 as the default first-pass renderer unless the user explicitly says not to use it. A request to create such a diagram authorizes the default image-generation stage; do not add a redundant pre-generation confirmation. After every generated image, show the raster draft and ask whether the user wants it reconstructed as editable SVG, vector PDF, editable PPTX, or a combination. Do not start reconstruction until the user agrees. Use the pure-SVG-first route only when the user explicitly rejects GPT Image 2 or explicitly asks for pure SVG/code-first output.
Classify the destination before drawing: paper mechanism figure, PPT/Poster graphic, or README/outreach infographic. A paper architecture request must expose the computation graph and pass references/paper-vs-presentation-diagrams.md; do not use slide-style stage cards, hero headings, callout dashboards, or presentation prose merely because they look polished. Presentation grammar is allowed only for an explicitly presentation-facing artifact and must be labeled accordingly.
Core Rule
Make every visual evidence-bearing, readable, and integrated with the manuscript. Start from a visual contract, not a template. Never invent data, numbers, statistics, baselines, sample sizes, architecture modules, data flows, training signals, labels, images, captions that imply unsupported results, or official venue rules.
Preserve strong user prompts. When an architecture prompt already specifies the scientific topology, components, labels, and visual intent, keep it verbatim and append only a compact content-fit refinement. Do not replace it with a longer house-style prompt. When the input is a manuscript or sparse notes instead, extract a factual diagram specification and build the prompt from that evidence. Treat prompt length as a cost and latency surface.
Use natural title case or sentence case for ordinary English displayed inside a figure, while preserving canonical uppercase for acronyms, initialisms, and standardized technical abbreviations. For example, use Visual Composer, GPT Image 2, Structure QA, and Editable SVG/PDF/PPTX, never VISUAL COMPOSER, Gpt Image 2, Structure Qa, or Editable Svg/Pdf/Pptx. Acronyms such as CCF, AI, GPT, QA, SVG, PDF, PPTX, and PNG remain uppercase. Do not use all caps for complete ordinary-language titles, module names, legends, axes, annotations, badges, or table headers merely for emphasis.