superwangkezheng-wq/openclaw-deep-research-master/deep-research-visuals

Deep Research Visuals

This skill is the visual routing layer for the Deep Research project. It replaces ad-hoc "figure-ready" prose with real, readable, source-traceable visual assets.

Core Policy

1. Source-first: inspect cited papers, PDFs, official pages, and web sources for reusable figures before drawing.
2. If a source figure is suitable and legally usable, save it with source URL, title, citation, and license/usage note.
3. If a source figure is absent, restricted, low-resolution, or not aligned with the argument, redraw an original figure.
4. Never ship only a figure-ready paragraph when the user asked for a figure, diagram, flowchart, or PPT visual.
5. Prefer PNG as the stable display artifact for Obsidian/PDF/PPT. Keep SVG/PDF/TIFF/script/source data beside it when useful.
6. Do not use draw.io foreignObject SVG as the only rendered artifact.

Composition, Not Either-Or

Read references/tool-routing.md when deciding tools.

This skill is an orchestrator. Do not treat tools as mutually exclusive. Combine them when the figure benefits from the intersection:

• nature-figure supplies the figure contract, scientific argument, panel hierarchy, restrained palette, typography, and publication/export discipline.
• Python/matplotlib/seaborn implements quantitative panels, matrices, roadmaps, radar/forest/interval charts, and final multi-panel assembly.
• draw.io supplies manually polished strategy/system panels that can be embedded into a larger Nature-style page.
• fireworks-tech-graph-ppt-chatgpt supplies native, fully editable PowerPoint architecture, data-flow, flowchart, sequence, Agent/RAG/memory, and tool-call diagrams when editable_technical_diagram=true. It remains a specialist panel owner under this orchestration layer.
• Mermaid/PlantUML/Graphviz produce reproducible structure, timeline, architecture, and evidence-map panels that can be rendered and then included as one panel in a composed figure.
• Manim/schemdraw/Bioicons produce specialist scientific, mathematical, electronic, or domain-icon subpanels.

The default question is not "A or B?" It is "which tool owns which panel, and what final composition makes the argument clearest?"

Default ownership choices:

• nature-figure: scientific, quantitative, evidence-bearing, roadmap, heatmap, radar, matrix, forest/interval, or multi-panel report/PPT figures. For Deep Research automation, use backend python unless the user explicitly asks for R.
• drawio: system frameworks, layered architecture, strategy maps, and diagrams that need manual executive polish.
• fireworks-tech-graph-ppt-chatgpt: technical diagrams that must remain native and editable in PowerPoint. Use deai-cn on Chinese labels before building, keep the generated .pptx, render and inspect a PNG preview, and do not substitute a screenshot or hand-authored OOXML.
• mermaid: simple flowcharts, timelines, sequence charts, and stage pipelines.
• plantuml: UML/C4/component/deployment diagrams.
• graphviz: DAGs, evidence maps, dependency/causal graphs, knowledge graphs.
• manim: mathematical or scientific concept frames.
• schemdraw: circuit/electronic schematics.
• bioicons: biology/chemistry icon support.
• diagrams: cloud/system infrastructure.

Nature-Figure Integration

For deep-research final delivery, the backend choice is explicit: backend=python. This avoids blocking a long-running final-delivery agent with "Python or R?" while still using the nature-figure contract.

Use the project visual Python runtime for all Python-backed rendering unless the caller explicitly overrides it:

${RESEARCH_VISUAL_TOOLS_ROOT:-${HOME}/.local/share/research-visual-tools}/venv/bin/python

Do not fall back to system python3 for matplotlib/nature-style rendering when the project venv exists. If the venv is missing or lacks packages, run the doctor and report the blocker instead of silently changing renderer.

Use nature-figure patterns for:

• technology roadmaps and density-equivalent curves
• evidence-strength bars or interval plots
• boundary/risk matrices and heatmaps
• business-fit radar or polar summaries
• multi-panel scientific concept pages
• semiconductor, AI accelerator, packaging, and platform comparison figures

Each nature-style generated figure should keep:

• editable SVG
• PDF
• stable PNG display artifact
• script under visual_assets/scripts/
• source data under visual_assets/data/ when quantitative

When another tool produces a subpanel, save that subpanel as SVG/PNG/PDF and record it in the same figure entry. For fireworks-tech-graph-ppt-chatgpt, keep the native .pptx as the editable artifact and the inspected PNG preview as the stable display artifact. The final assembled figure can still be deep-research-visuals:nature-figure+drawio, deep-research-visuals:fireworks-tech-graph-ppt-chatgpt, deep-research-visuals:nature-figure+graphviz, or another explicit combination.

Required Output

visual_asset_plan.json must include one entry per figure:

{
  "visual_asset_policy_version": "2026-06-01",
  "source_first": true,
  "figures": [
    {
      "figure_id": "F1",
      "title": "",
      "purpose": "",
      "figure_contract": {
        "core_conclusion": "",
        "archetype": "quantitative grid | schematic-led composite | image plate + quant | asymmetric mixed-modality figure | system diagram",
        "evidence_chain": []
      },
      "source_search": {
        "performed": true,
        "queries": [],
        "candidate_sources": []
      },
      "decision": "reuse_source_figure | draw_original",
      "tool": "deep-research-visuals:nature-figure | fireworks-tech-graph-ppt-chatgpt | drawio | mermaid | plantuml | graphviz | manim | diagrams | schemdraw | bioicons",
      "toolchain": ["nature-figure", "matplotlib"],
      "panel_sources": [
        {
          "panel_id": "a",
          "tool": "nature-figure",
          "artifact": "visual_assets/F1_panel_a.svg",
          "role": "hero quantitative panel"
        }
      ],
      "backend": "python | r | drawio | mermaid | plantuml | graphviz | manim | diagrams | schemdraw | source",
      "text_skill": "deai-cn | not_applicable",
      "editable_artifact": "visual_assets/scripts/F1.py",
      "rendered_artifact": "visual_assets/F1.png",
      "svg_artifact": "visual_assets/F1.svg",
      "pdf_artifact": "visual_assets/F1.pdf",
      "source_url": "",
      "license_or_usage_note": "",
      "qa_status": "opened_or_rasterized",
      "status": "reused_source_figure | drawn_rendered"
    }
  ]
}

visual_asset_log.md must explain source inspection, reuse/redraw decisions, tool choice, QA result, and the final display artifact referenced by final_delivery.md / ppt_outline.md.

Huawei Tao Law Defaults

For huawei-tao-law-20250607-style research, prefer:

• source timeline: nature-figure if evidence-tiered, otherwise Mermaid
• tau concept frame: nature-figure or Manim
• density-equivalent vs lithography-node comparison: nature-figure
• applicability boundary matrix: nature-figure
• industry ecosystem: draw.io or Graphviz
• Lenovo business mapping / AI PC route: nature-figure when quantitative, draw.io only when it is a strategy architecture map

Verification

Run:

zsh skills/deep-research-visuals/scripts/deep-research-visuals-doctor.sh

Before marking final delivery ready, rendered assets must be nonblank, readable, and referenced by the final Markdown/PPT outline.

When editable_technical_diagram=true, also require the specialist skill's geometry checks to contain no ERROR: output and visually inspect its PNG preview. Record fireworks-tech-graph-ppt-chatgpt in toolchain and panel_sources, deai-cn in text_skill, the native .pptx in editable_artifact, the preview PNG in rendered_artifact, and qa_status=preview_inspected.