Academic Draft and Paper Analyzer (Figure Planner)

Produce a human-readable figure strategy and a machine-readable Figure Plan v1. Plan figures around the paper's claims and reader questions, not around a fixed count or a generic pipeline template.

Read references/missing-info-policy.md when the paper is incomplete. If a repository handoff or extracted reference-style profile exists, carry it forward without renaming fields.

Dual-Track Planning Workflow

The analyzer operates in one of two modes depending on the input:

1. Draft-to-Figure Fast-Track (草稿敏捷直出):
   • Trigger: Input is notes, an outline, or an early draft (Markdown, text, or draft sections) without complete experimental/analysis results.
   • Target: Focus exclusively on Figure 1: Overall Framework / Methodology Overview.
   • Policy: Do NOT force a multi-figure plan (ablation/data behavior). Do NOT warn about missing experimental/analysis sections.
2. Camera-Ready Multi-Figure Plan (完整定稿多图规划):
   • Trigger: Input is a complete manuscript with experimental results (Markdown, LaTeX, PDF, or full text).
   • Target: Systematically plan the multi-figure suite with claim verification and publication constraints.

Input contract

• Prefer: manuscript text or source, abstract, method, experiments, target venue/page limit, semantic architecture handoff, and any reference figures.
• Accept: Markdown (.md) draft or notes, outline, title plus abstract, local PDF, paper URL/HTML, Word/LaTeX.
• A URL is a paper source only after its content is inspected; do not classify every URL as a code repository.
• If a PDF or webpage cannot be read in the current environment, report that limitation rather than inventing paper structure.

Output contract

Include:

1. Paper overview: question, contribution, evidence, and intended venue constraints.
2. Completeness statement: sections and artifacts actually inspected.
3. Per-figure recommendation with a controlled type and must, strong, or nice priority (in Draft mode, emit exactly one primary Figure 1).
4. A one-sentence communication goal: what the reader should understand after viewing the figure.
5. Required nodes, edges, authority boundaries, and forbidden implications.
6. Aspect ratio and final publication width hint.
7. Figure Plan v1 JSON.

Controlled figure types: Overall Framework, Network Architecture, Module Detail, Comparison/Ablation, Data Behavior, Concept/Motivation, Protocol/Sequence, Timeline/Lifecycle.

Workflow

1. Identify Input Mode and Structure

• In Draft Mode (.md notes / outline / unstructured prose / draft):
  • Structured Markdown:
    • Headers (#, ##): Primary semantic container zones (e.g. Input Context, Core Framework, Optimization Objective).
    • Lists (1. 2. 3., - ) and arrows (->): Sequential stages, data flow, and pipeline steps.
    • Bold terms (**Name**): Canonical node/module labels (visible_text).
    • Inline code (`B x C x H x W`): Tensor dimensions or data descriptions.
  • Unstructured / Plain Prose Notes:
    • Extract entities, transforms, and dataflow by tracing the grammatical subject-verb-object sequence (e.g. "Input X is encoded by Y and supervised by loss Z").
    • Group into 3 canonical panels: Input / Context -> Core Method / Interaction -> Output / Supervision.
  • Skip formal claim verification and proceed straight to narrative topology for Figure 1.
• In Full Paper Mode:
  • Map Introduction, Method, Experiments, Analysis, and Limitations. For each claimed contribution, record the source span and the evidence that could support a visual statement.

2. Assign Visual Jobs

Recommend a figure only when a visual materially improves understanding:

Reader question / Objective	Common type	Typical priority
What is the end-to-end idea and authority/data flow?	Overall Framework	must (Default for Drafts)
What is the conceptual motivation or problem setting?	Concept/Motivation	strong or must
What is the internal executable structure?	Network Architecture	must or strong
How does the central mechanism work?	Module Detail	must or strong
Which choices matter empirically?	Comparison/Ablation	strong
How does behavior change over data, time, or conditions?	Data Behavior	strong or nice
How do components interact over time or protocol?	Protocol/Sequence / Timeline	strong

Do not use venue stereotypes or fixed counts as requirements. For draft notes, emit a single high-impact Figure 1.

3. Design the narrative topology

For each figure specify:

• communication_goal and claim_scope;
• hero_element: the dominant visual story, not merely the largest box;
• required_nodes and required_connections from source evidence or draft structure;
• secondary_context that may be dropped under space pressure;
• forbidden_claims and forbidden_connections;
• authority_boundaries when agents, tools, or external oracles are involved;
• text_budget: short labels, with detail reserved for the caption;
• reference_style_profile if the user supplied a reference image.

An Overall Framework need not be a left-to-right chain. Choose among a loop, storyboard, asymmetric modular collage, layered authority diagram, central mechanism with satellites, or pipeline according to the scientific story.

4. Set publication geometry

• Overall Framework: usually 16:9 or 3:2 at double-column width (183 mm).
• Network Architecture / Pipeline: 16:9, 3:2, or tall layout when topology requires it.
• Module Detail / Concept: commonly 4:3 or 1:1.
• Comparison/Ablation: match the number and reading order of panels.
• Data Behavior / Timeline: let axes and sequence determine geometry.

5. Emit Figure Plan v1

{
  "schema": "academic-figure/FigurePlan@1",
  "source_revision": "<paper-or-repo-revision>",
  "venue": null,
  "sources": [],
  "figures": [
    {
      "figure_id": "fig1",
      "figure_type": "Overall Framework",
      "priority": "must",
      "communication_goal": "<one sentence>",
      "claim_scope": ["<evidence-backed claim or draft objective>"],
      "hero_element": "<loop|mechanism|modular collage|pipeline|other>",
      "required_nodes": ["<component-id>"],
      "required_connections": ["<from-id> -> <to-id>: <kind>"],
      "authority_boundaries": [],
      "secondary_context": [],
      "forbidden_claims": [],
      "forbidden_connections": [],
      "aspect_ratio": "16:9",
      "final_width_mm": 183,
      "style_profile_hint": null,
      "reference_assets": [],
      "open_questions": [],
      "confidence": "high|partial|sparse",
      "review_status": "pending|confirmed|waived"
    }
  ]
}

The example defines fields only. Replace every placeholder with sourced content or an explicit null/empty value.

Sparse input & Draft Handling

• Single .md draft or notes: Extract pipeline directly from headers and bullet points; plan Figure 1 immediately.
• Title and abstract only: plan high-level visual jobs; do not invent submodules.
• Method without experiments: plan method framework; do not block or raise spurious errors.
• Repository handoff only: produce a system-centric plan and flag narrative review.
• Reference image only: analyze visual grammar, but request or locate the target system content before planning its topology.

Stop

Stop after delivering the strategy and Figure Plan v1. Do not generate prompts or images unless the user requested the downstream workflow.