Engineering Figure Agent

Engineering Figure Agent

Use this skill after the figure goal is already reasonably clear. For conceptual framework or experiment-design figures, prefer conceptual-figure-workflow first so academic-research can review the theory structure and produce a figure brief before this skill handles schematic or image execution.

Boundary

Good fit:

• Turn a figure brief into a conceptual diagram, workflow figure, schematic, graphical abstract, redraw, or edited image.
• Produce an editable .drawio figure when labels, arrows, topology, or later manual revision must remain exact.
• Build prompts, generate conceptual panels, and refine existing briefs.
• Assemble already rendered quantitative panels with a schematic when the panel contract is fixed.

Not the main tool for:

• Deciding from scratch what claim the paper should visualize.
• Full paper-argument or reviewer-style critique.
• Reviewing theoretical structure, variable relationships, mechanisms, hypotheses, or claim-evidence boundaries for a conceptual framework.
• Creating empirical, statistical, benchmark, ablation, map, or multi-panel results figures from numeric data. Use nature-figure for those tasks.

If the user is still deciding the figure claim, panel logic, or caption argument, use conceptual-figure-workflow or academic-research first.

Legacy plot scripts and references remain for upstream compatibility. Do not load or invoke them in normal routing; send numeric-data figure work to nature-figure.

Core Decision

• openai-image mode: illustrative conceptual figures, graphical abstracts, reference-inspired redraws, and image edits where visual synthesis matters more than object-level editability.
• drawio mode: label-heavy conceptual frameworks, mechanism diagrams, architecture/workflow figures, or any figure requiring exact nodes, arrows, topology, and editable vector-like objects. Route execution to the installed drawio skill while preserving this skill's reviewed figure contract.
• mixed mode: use nature-figure to produce quantitative panels, then use this skill only for conceptual panels or final schematic-led composition. Use Draw.io when the final composition must remain editable.

Never use image generation for exact values, axes, or benchmark geometry.

Default Workflow

1. Inspect the user input and decide whether a figure brief is already present.
2. If needed, create a brief using references/figure-brief-spec.md.
3. Lock the semantic contract before rendering: nodes, directed relationships, evidence status, visible-text allowlist, prohibited implications, and any unresolved issue ledger.
4. Choose openai-image, drawio, or mixed mode. Prefer Draw.io when a mislabeled arrow or generated-text error would change the scientific claim.
5. For image execution, prefer the prompt-builder scripts; read template references only when template wording needs customization.
6. For Draw.io execution, hand the semantic and render contracts to drawio. Keep the .drawio source as the primary editable artifact.
7. For a mixed figure, obtain quantitative panels from nature-figure without asking image generation to redraw their values, axes, or geometry.
8. When the brief selects explore, create 2-4 structurally distinct low-cost candidates, compare them against one issue ledger, then refine only 1-2 directions. Do not use candidate exploration by default.
9. Audit the final figure against the semantic contract. Reject unsupported arrows, unapproved visible text, reversed flows, duplicated concepts, and decorative elements that imply evidence not present in the brief.
10. Save prompt/spec/source/output paths when files are produced, then run only the checks needed for the touched path.

Reference Loading

Read only what is needed:

• references/figure-brief-spec.md: brief structure and mode rules.
• references/semantic-render-audit.md: semantic graph, render graph, text allowlist, candidate comparison, and final audit.
• references/publication-figure-design.md: publication styling defaults.
• references/provider-selection.md: OpenAI-first setup and optional Gemini/Banana compatibility.
• references/highres-policy.md: high-resolution and fail-closed rules.
• references/chinese-labels.md: Chinese label readability rules.
• references/engineering-figure-templates.md and references/materials-science-figure-template.md: template wording when script output needs customization.
• references/editable-figure-handoff.md: Draw.io-first editable handoff and optional SVG reconstruction.

Scripts

• Build a prompt without network calls:

python scripts/build_engineering_figure_prompt.py --figure-template system-architecture --lang en "technical background"
python scripts/build_materials_figure_prompt.py --materials-figure mechanism-figure --lang en "scientific background"

• Generate or edit a conceptual image:

python scripts/generate_image.py --provider openai --figure-template system-architecture --lang en "technical background"
python scripts/generate_image.py --provider openai --materials-figure mechanism-figure --lang en "scientific background"

• Use the unified CLI:

python scripts/efa.py prompt --figure-template system-architecture "technical background"
python scripts/efa.py prompt --materials-figure mechanism-figure "scientific background"
python scripts/efa.py check

For the meaning of approved nodes and edges, read semantic approval scope. Directly supported semantics do not require an extra approval round; new or conflicting scientific assertions still require a user decision.

Quality Rules

• Do not fabricate measurements, benchmark values, hardware specs, or unsupported causal claims.
• Keep arrows, reading order, and module hierarchy explicit.
• Treat every arrow as a scientific assertion. Match it to an approved semantic edge and preserve direction, relation type, and certainty.
• Render only allowlisted visible text. Do not let image generation invent acronyms, metrics, stages, captions, or explanatory labels.
• Favor white backgrounds, publication-style spacing, and readable labels.
• Use prompt-first workflows when layout fidelity matters.
• Use the official OpenAI image path by default. Use Gemini, Banana, relays, or other providers only when the user explicitly selects them.
• For high-resolution requests, follow references/highres-policy.md.
• For third-party providers, require explicit opt-in before sending keys or user files.
• For Chinese figures, preserve standard technical notation when it improves clarity.