Scientific Image Prompting

Scientific Image Prompting

Purpose

This skill is for scientific illustrations, not data plots.

Use it to turn a research intent into a prompt package that is:

• publication-aware
• visually precise
• traceable
• explicit about what is conceptual vs. measured

Hard Routing Rule

Before writing any prompt, classify the request into exactly one route:

1. data_figure

• Real experiment or statistical figure
• Examples: ROC, PR, heatmap, volcano, PCA, UMAP, histogram, line chart, confusion matrix
• Action: stop using this skill and route to experiment-lab, nature-figure, or validated Python/R plotting

2. deterministic_diagram

• Flowchart, architecture, workflow, study design, pipeline, mechanism path that should be clean and diagrammatic
• Action: prefer fireworks-tech-graph or another deterministic diagram workflow

3. ai_scientific_illustration

• Graphical abstract, concept art, mechanism imagination, cover art, non-quantitative scientific explainer
• Action: continue with this skill, then hand off to image-generation

If the request contains measured values, axes, significance claims, or looks like a result figure, it is not ai_scientific_illustration.

Allowed Figure Types

Use one of these figure_type values:

• graphical_abstract
• mechanism_illustration
• workflow_schematic
• study_design
• cover_art
• concept_explainer

Prompt Contract

Create prompt.json with this contract:

{
  "prompt_contract_version": "scientific-image-prompting.v1",
  "route": "ai_scientific_illustration",
  "figure_type": "graphical_abstract",
  "scientific_goal": "What the figure should explain",
  "must_include": [
    "Required scientific entities, stages, organs, cells, devices, molecules, or scene elements"
  ],
  "must_not_invent": [
    "Any measured result, axis, p-value, or unsupported biological / technical claim"
  ],
  "label_strategy": "short labels only | no embedded labels | leave whitespace for post-edit annotation",
  "composition": "panel structure, focal path, camera angle, negative space",
  "style": "flat vector-like | polished 3D editorial | biomedical infographic | clean concept art",
  "lighting": "if applicable",
  "color_palette": "3-5 colors with scientific publishing intent",
  "reference_requirements": [
    "what reference images are needed and why"
  ],
  "prompt": "Final English generation prompt",
  "negative_prompt": "What must be excluded",
  "technical": {
    "aspect_ratio": "16:9",
    "image_size": "4K",
    "output_mime_type": "image/png",
    "scientific_mode": true
  }
}

Writing Rules

• Always write the final generation prompt in English.
• Chinese may be used only for local explanation to the user.
• Keep the figure conceptual unless the user supplied real measured content to place into a non-plot illustration.
• Prefer white or very light backgrounds for paper-ready figures.
• Minimize embedded text inside the image. If labels are needed, keep them short and publication-like.
• If visual fidelity matters, gather reference images first.

Scientific Guardrails

• Do not fabricate data-like figures.
• Do not generate fake microscopy, fake western blots, fake sequencing plots, fake statistical charts, or fake benchmark panels as if they were results.
• Do not imply that an imagined mechanism has been experimentally validated unless the user explicitly provided that evidence.
• If the figure is conceptual, ensure the downstream deliverables say so.

Required Deliverables

For scientific illustration requests, prepare these files in outputs:

• scientific-illustration-4k.png
• prompt.json
• prompt_audit.md
• caption.md
• ai_disclosure.md

Audit Notes

prompt_audit.md must briefly record:

• chosen route
• chosen figure_type
• why AIGC is appropriate here
• what was intentionally excluded to avoid fake data presentation
• provider/model choice, including whether it came from the active Settings configuration or an explicit override
• size/output target: 4K, PNG

ai_disclosure.md must explicitly state that the image is a conceptual or illustrative figure generated with AI assistance and should not be interpreted as raw experimental evidence.

Handoff to Image Generation

After prompt.json is ready, call image-generation in scientific mode with:

python /mnt/skills/public/image-generation/scripts/generate.py \
  --prompt-file /mnt/user-data/outputs/prompt.json \
  --output-file /mnt/user-data/outputs/scientific-illustration-4k.png \
  --manifest-file /mnt/user-data/outputs/generation_manifest.json \
  --aspect-ratio 16:9 \
  --scientific-mode \
  --image-size 4K \
  --output-mime-type image/png

Default to the active image provider/model configured in Settings. Only add --provider, --model, or --base-url when the user explicitly wants to override the configured provider.

If the user asked for a lower-cost draft, use:

• --draft-mode
• or explicitly --model gemini-2.5-flash-image

Final Check

Before delivery, confirm:

• this is not a disguised data figure
• the image is conceptual and paper-appropriate
• prompt, manifest, caption, and disclosure files all exist
• the final output is PNG and intended as 4K