FigFox-Gen-skill

Core workflow

Run exactly one image-generation pass and stop at PNG1:

Methodology + optional reference
  -> Context 1: recurring domain visual language
  -> Context 2: content-to-visual plan
  -> Context 3: inspected FigureBench construction evidence + selected palette group
  -> Creative Director prompt: brief + targeted scholarly paper-SVG crops
  -> Prompt 1 bundle
  -> direct image generation
  -> final PNG1
  -> stop

There is no post-generation conversion, temporary render, repair loop, or second image. The local helpers validate files, provenance, and prompt contracts; they do not observe model calls or guarantee scientific correctness.

Canonical visual examples

When the user cites the bundled examples, treat these exact files as a style canon:

• assets/generated-figures/02-latent-diffusion.png
• assets/generated-figures/01-figfox-gen-workflow.png
• assets/generated-figures/01-figfox-gen-workflow-zh.png
• assets/generated-figures/03-musicot.png
• assets/generated-figures/04-alphafold3.png

Inspect the closest example before choosing a visual language. Borrow its restrained composition, alignment, spacing, flat construction, readable labels, and plain arrows; do not merge unrelated examples into a collage or invent a more decorative aesthetic. These examples are visual references, not active palette sources, and they never override scientific meaning or explicit user constraints.

Start a run with input/methodology.md and, when supplied, input/user-reference.<ext>. Use the exact JSON fields in artifact schemas. Before work begins, run:

python scripts/figure_workflow.py check-installation --root .

1. Context 1: recurring domain visual conventions

Read the Methodology with the language model and extract the scientific domain, core topic, mainline, likely figure class, named concepts, and explicit user constraints. Search for 3–4 scholarly papers in that domain. Prefer arXiv and native SVG/HTML figures; otherwise use credible papers with clearly extractable panels. Inspect actual figure regions, not titles or abstracts alone.

Use the first representative paper figure selected during this domain search as the colour-count anchor. Inspect its visible panel, record its dominant-colour count as dominant_colour_count (a positive integer supported by an eligible palette group), and carry that count into Context 3. Use the remaining 2–3 papers to corroborate the anchor and the visual conventions; do not replace the anchored count with a subjective guess or copy source colours. A broad field may justify a larger count only when the representative figure visibly uses more dominant roles and the retained papers support that reading. Record source URLs and crop paths in references/web/manifest.json and references/web/crops/. Every mapped crop states its target component, what to borrow, what must change, and why the result remains human-editable.

If fewer than three credible figures remain, continue searching or report insufficient evidence. Do not invent recurrence.

python scripts/figure_workflow.py validate-context --run RUN --context 1

2. Context 2: content-to-visual plan

Combine the Methodology, Context 1, and optional user reference. The user reference has strong authority for structure, layout, emphasis, and visibly human-made basic visual treatment. When a user reference is attached or explicitly cited, enter a reference-fidelity lock: treat it as the canonical visual specification, match its composition, spacing, hierarchy, and visual grammar, and preserve its line weight, corner-radius, fill treatment, typography scale, arrow grammar, and sample treatment wherever the Methodology permits. Change only labels, scientific content, and geometry required by the Methodology. Do not beautify, complicate, stylize, recompose, or switch to a different visual language because the model prefers it. Ignore only generated-looking, fake, or decorative parts, and never take active palette colours from it.

Express the scientific mainline in text, then map every block, structure, relationship, and required label to an exact visual treatment in context/context-2-content-visual-plan.json. In-image text is limited to concise block names, structure names, necessary scientific labels, terms, and relationships; explanatory prose stays in planning.

Normal treatments are human-producible basic geometry, recurring domain-paper constructions, deliberate manual/stylus drawings, draw.io-like editable structures, or real photographic crops when scientifically necessary. Mark any other treatment as special and explain whether a human would construct it geometrically, draw it by hand, or obtain it as a real photo. Every visual must have a semantic role and construction provenance.

python scripts/figure_workflow.py validate-context --run RUN --context 2

3. Context 3: construction evidence, palette groups, and taste

Read FigureBench visual selection and taste rules.

The installed FigureBench pack contains exactly 30 complete development images. Inspect at least two distinct complete references, then continue adaptively until every Context 2 need has credible evidence: shapes, frame/container families, connectors, layout relationships, and special visualizations. There is no fixed target or maximum image count. Stop only at complete coverage.

Complete references are inspection sources, never unexplained Prompt 1 attachments. For every useful region, preserve normalized coordinates and a component-specific contract in references/figurebench/crops/request.json. The contract names what to borrow, what must change so the result is a variant, and why the treatment is human editable. Materialize and validate the crops:

python scripts/figure_workflow.py rank-references --run RUN
python scripts/figure_workflow.py crop-references --run RUN
python scripts/figure_workflow.py validate-reference-coverage --run RUN

FigureBench supplies geometry, layout, spacing, connectors, and human-edited finish. It never supplies active palette colours.

The local references/palette-library.json stores named palette groups; each group contains several role-labelled colours, not one colour. At the start of each run, randomly select one eligible named group and record its id; use multiple colours from that group as needed. Use select-palette --run RUN for an unseeded selection or select-palette --run RUN --seed SEED when the run must be reproducible. This is a palette-group lineage rule, not a monochrome or single-colour rule. If the selected group lacks a required functional role, only an evidenced tint, shade, tone, analogous neighbour, compatible neutral, or controlled contrast may extend it. Never mix a second library group or take colours from the user reference, FigureBench, or domain papers. Taste is a low-priority soft constraint for spacing, hierarchy, rhythm, balance, restraint, and human-edited finish. The final figure must use exactly Context 1's observed dominant-colour count from the selected group. Other swatches are subordinate neutral, tint, shade, or support roles and must not be promoted beyond that anchored count. If a group cannot provide enough roles, select another eligible group or report that no palette group can support the evidence.

Write context/context-3-visual-kit.json from the materialized crops, coverage matrix, selected palette-group lineage, and taste constraints, then validate:

python scripts/figure_workflow.py validate-palette --run RUN
python scripts/figure_workflow.py validate-context --run RUN --context 3

4. Creative Director: pre-PNG1 visual ideation

The Creative Director runs once after Contexts 1–3 and before PNG1. It is a bounded ideation pass, not an image-generation pass. It may propose a concrete, scientifically relevant treatment for a planned component, but it must not redraw the whole figure or invent decorative assets.

Before choosing a style, the Creative Director must model how a human would actually build the figure in an editor. Use this construction order as a hard planning sequence:

1. Choose the base: establish the main canvas, containers, and simple geometry first (usually rectangles or deliberately adjusted rounded rectangles). A subtle fill variation is allowed only when it belongs to the base shape and remains editable; never use a gradient to hide an unplanned structure.
2. Build the content on the base: add the planned structures and modules. Use the simplest geometry that carries the meaning. If the method needs a topology, grid, model block, or other known construction, look for a real scholarly SVG or extractable figure and reuse a targeted crop as evidence instead of inventing a fake topology. If it needs an input sample, use a real or explicitly documented sample-like crop (text, image, audio, or data), not generic placeholder lines.
3. Add restrained arrows: draw plain, readable connectors after the objects are placed. Arrowheads and paths must express direction and relation; do not make arrows decorative, glossy, multi-coloured, or needlessly curved.
4. Place exact text: add concise block names, terms, and relationship labels after the structure is stable. Never replace real labels with repeated horizontal filler lines.
5. Place the visual below/next to its label: keep each explanatory visual close to the text it explains. Prefer a real sample, a paper-derived editable crop, a regular copied geometry, or a library symbol with clear provenance. A person, warning, chat symbol, or other special object must follow its human source and semantic role; it may not become a pasted sticker.

Construction invariants: a regular grid or copied geometry must have consistent spacing and no accidental missing cells or broken lines; a geometric block uses a flat fill or one controlled fill per block, never multiple arbitrary colours inside one block; noise is shown with deliberate repeated points or a real noise image only when the Methodology calls for it; photographs are real photographs when a photo is scientifically necessary. These rules are planning gates, not post-hoc taste advice.

Compile its prompt:

python scripts/figure_workflow.py build-creative-director-prompt --run RUN

The model returns creative-director/brief.json. Validate it:

python scripts/figure_workflow.py validate-creative-director --run RUN

If a new idea needs a construction not already covered by Contexts 1–3, the Creative Director must locate a real scholarly paper figure available as SVG or extractable SVG/HTML, inspect its pixels, and request only a targeted crop under references/web/crops/creative-director/. Each crop must include the target component, HTTPS source_url and evidence_url, source_format: "svg", nonempty borrow and must_change lists, and a human-editability reason. Never invent a source, attach a complete paper figure, copy its labels or palette, or use a sticker-like cutout. If no new treatment is needed, return an explicit no_external_svg_needed brief with no crop. Palette lineage and all Prompt 1 anti-AI constraints remain in force.

5. Prompt 1 and final PNG1

Compile Prompt 1 only after Contexts 1–3, the Creative Director brief, and every mapped crop file exist:

python scripts/figure_workflow.py build-prompt1 --run RUN

The bundle contains the Methodology, Contexts 1–3, the Creative Director brief, the optional reference, mapped scholarly crops, every mapped FigureBench crop, and any Creative Director paper-SVG crop. A crop guides only its declared component and its borrow/must_change contract. It cannot donate source labels, source colours, source proportions, or a complete composition.

Prompt 1 contains exact mainline, block names, relationships, content-to-visual mapping, crop contracts, selected palette-group lineage, layout/taste constraints, concise labels, and anti-AI invariants. Enforce these defaults:

• when a user reference is attached or cited, the reference-fidelity lock is hard: match its composition, spacing, hierarchy, and visual grammar before making any stylistic choice; preserve its human-made construction cues and change only what the Methodology requires;

• do not beautify, complicate, stylize, recompose, or switch visual language merely because the image model prefers a different look; reject any output that drifts from the reference without a scientific reason;

• no meaningless dots, tiles, floating symbols, purposeless boxes, irrelevant ornament, unjustified extreme contrast, decorative gradients, glow, decorative shadow, fake cartoons, or shapes without human construction provenance;

• no default numbered 1/2/3/4 planning labels and no generic blue title-strip/content-box cards unless the Methodology explicitly requires them;

• never box off the upper portion of a module with a horizontal divider and centered title;

• never paste a sticker-like cutout, clip-art badge, medal, seal, or raster badge;

• every visual must be semantically related to its text and editable by a human.

• follow the human construction order: base geometry first, then meaningful content, plain arrows, concise text, and a nearby explanatory visual; do not reverse this into decorative cards with filler content;

• use real samples or targeted scholarly-paper construction crops for inputs, topologies, grids, and model diagrams when they exist; never invent a fake topology or pseudo-sample when a faithful human-editable source can be found;

• keep grids and repeated geometry exact and regular, keep each geometric block to a flat or single controlled fill, except that a single base/container may use a subtle deliberate fill transition when it is part of the planned geometry; never use gradients inside a block or as decorative polish. Keep arrows visually subordinate to the objects and text.

The first and only image-generation pass receives prompt-1/prompt.md and all manifest attachments. Save the complete labelled result as png1.png. PNG1 is the final deliverable for this workflow. Do not convert it to SVG, render a temporary derivative, diagnose a later pass, or generate another image.

6. Completion and deterministic validation

Write run-manifest.json with the canonical paths in artifact schemas, then run:

python scripts/figure_workflow.py validate-run --run RUN

A valid run contains exactly one image artifact, png1.png; one Prompt 1 root; the Creative Director prompt and brief; the three Contexts; the scholarly web manifest; the FigureBench request and materialized crop manifest; and the final Prompt 1 attachment manifest. The validator checks path safety, provenance, palette lineage, crop replay, Creative Director source contracts, prompt determinism, and PNG validity. It does not prove image-generation call counts or scientific truth. Authors must inspect PNG1 before publication.