Research Draw

Research Draw

Create or use a high-quality scientific illustration reference first, then redraw it as an editable draw.io diagram. The raster image is a composition and geometry guide, not the final source of truth.

Required Skill Sequence

1. Use imagegen for the reference-image stage only when no suitable raster reference already exists.
2. Use research-drawio-skill for the draw.io tracing stage.
3. During the draw.io stage, use research-drawio-skill for layout, XML, export, and QA, but override its primitive-glyph default for complex recognizable elements with this skill's SVG glyph workflow.
4. Do not embed the generated bitmap as the final diagram body unless the user explicitly asks for a raster-backed figure.

Routing Protocol

1. Read references/two-stage-workflow.md before starting.
2. Read references/imagegen-reference-stage.md before generating the raster reference image.
3. If tracing an existing PNG/JPG/WebP, read references/png-layout-extraction.md before authoring or editing .drawio.
4. Read references/drawio-tracing-stage.md before authoring or editing .drawio.
5. If complex real-world elements are present, such as animals, organs, instruments, cells with detailed morphology, or domain icons, read references/complex-asset-sourcing.md. Complex elements must use the SVG glyph workflow there, not draw.io primitive assembly.
6. Before final delivery, read references/consistency-loop.md and run the loop. For GPT/imagegen/raster tracing where close imitation matters, this is a minimum three-round strict pixel-level export/compare/fix loop.
7. After the final strict pixel comparison, run the final text completeness audit in references/consistency-loop.md before claiming delivery.
8. For GPT/imagegen/raster tracing where close imitation matters, make the draw.io stage load research-drawio-skill strict visual comparison rules.
9. If formulas appear, make the draw.io stage load the math reference from research-drawio-skill.
10. If online SVG assets are needed during the draw.io stage, use the local add-svg skill and follow its source gate. If self-designed SVG glyphs are sufficient, author them directly under the complex asset workflow.

Operating Rules

• Establish the scientific message before image generation.
• Keep the reference image label-light. Final labels, formulas, and captions should be rebuilt as editable draw.io text.
• Treat imagegen text as unreliable unless it is visually inspected and correct.
• Use imagegen to explore composition, visual hierarchy, spatial grouping, lighting-free style, and object silhouettes.
• When a PNG/JPG/WebP reference exists, trace geometry before redesigning: preserve relative object positions, sizes, gutters, and visual density unless the user asks for redesign.
• Use draw.io primitives for abstract structure: containers, network blocks, arrows, simple matrices, simple charts, and editable labels.
• Do not assemble complex recognizable objects from draw.io primitives by default. Animals, organs, detailed cells, instruments, anatomy, protein cartoons, and similarly silhouette-sensitive objects should become dedicated SVG glyphs.
• For every complex SVG glyph, render it at the target size and compare it against the corresponding reference crop before insertion. Iterate the SVG until the silhouette, scale, and color family are close enough, or record the remaining strict-comparison failure.
• Use draw.io primitives for a complex object only when the object is actually simple/abstract, no reference crop exists, or the user explicitly prioritizes full draw.io editability over visual fidelity.
• Preserve scientific correctness, but do not use editability, vectorization, formula rebuilding, or scientific-cleanup language to excuse avoidable visual mismatch. If strict comparison fails, keep editing or report the task as not yet passed.
• Run the consistency loop after tracing. Fix blocking geometry, semantic, and draw.io QA mismatches before delivery, or state the remaining mismatches.
• When the goal is close imitation of a GPT/imagegen reference, export the draw.io PNG and run quantitative comparison in at least three iterations. Use the worst mismatch tiles and diff overlays to drive each additional edit pass.
• After pixel comparison passes or stops, perform a text completeness audit: compare intended labels, legends, panel letters, formulas, axis text, and annotations against the draw.io text cells and exported preview. Fix missing, truncated, misspelled, duplicated, or pseudo-text remnants before delivery.
• Do not claim completion with wording such as "strict pixel comparison did not pass because this is an editable redraw" or "metrics are documented in trace notes." A failed strict comparison remains a failed QA gate unless the user explicitly accepts a different target.

Output Contract

Deliver or prepare:

reference image: path to generated raster concept
editable source: .drawio file
trace notes: mapping from reference regions to draw.io modules, including
  approximate reference coordinates for major regions
asset notes: searched/selected assets, license notes, and self-designed fallbacks
svg glyphs: paths to authored/selected SVGs for complex objects, with insertion
  bbox and data-URI encoding notes
glyph comparison: per-complex-element reference crop, rendered SVG PNG,
  metrics/diff paths, pass/fail, and remaining mismatches
consistency loop: at least three iteration records for close tracing, fixes
  made, remaining mismatches
strict comparison: per-iteration metrics, pass/fail, diff overlay paths, worst
  mismatch tiles, and final status as passed or not passed
text completeness audit: source/intended text inventory, draw.io text inventory,
  missing or changed labels/formulas, fixes made, final pass/fail
QA: draw.io validation result and remaining risks

If image generation is unavailable, skip the raster stage only after stating the blocker, then proceed with research-drawio-skill directly.