Analyze repositories, papers or drafts, and reference images into handoff-ready results. This skill does not write prompts or generate images.
Check the input content first, then choose a path. Do not treat every URL as a repository: after opening the page or body text, determine whether it is a paper, a code repository, or an image.
Check these three things:
- Body text, sections, or LaTeX/Markdown structure → paper or draft
- Source code, README, dependencies, and entry scripts → repository
- The image itself, figure captions, or embedded images in a PDF → reference image
Routing
Draft, outline, or complete manuscript
Markdown, LaTeX, PDF body text, or the body text after a paper URL is opened, read references/paper.md, produce Figure Plan v1.
- Notes, outlines, early drafts: plan only Figure 1 (Overall Framework); do not force in experiment figures.
- Complete manuscript with experimental results: plan multiple figures according to the paper's claims.
- When the current environment cannot read a PDF or web page, report the limitation; do not fabricate section structure.
Repository
A local repository path, or a URL confirmed to be a code repository after opening, read references/repo.md, produce Semantic Architecture Handoff v1. When the task or framework classification is uncertain, additionally read references/keywords.md。
Paper plus repository
The paper and user intent determine the narrative and topology. The repository only verifies parameters, tensor dimensions, module names, and execution direction; it does not introduce a separate architecture.
Reference image
External style reference images, figures in a PDF, or images, read references/reference-figure.md, produce ReferenceAnalysis v1.
- Use the PDF reader for PDF body text and figure captions.
- Only when image candidates need to be extracted, run
scripts/extract_pdf_figures.py。 - This script only filters candidates by size; it is not a structure classifier. Keep, discard, or uncertain is still determined by evidence.
When evidence is insufficient, read references/missing-info-policy.md. Mark missing items; do not fabricate modules, edges, or values.
Handoff
Fields and JSON shapes follow the definitions in the corresponding reference; the schema is not rewritten here:
- Semantic Architecture Handoff v1（
academic-figure/SemanticArchitecture@1）→references/repo.md - Figure Plan v1（
academic-figure/FigurePlan@1）→references/paper.md - ReferenceAnalysis v1（
academic-figure/ReferenceAnalysis@1）→references/reference-figure.md
When multiple inputs appear at the same time, each follows its own reference, then merge according to 'the paper determines the narrative, the repository only verifies'. Do not change field names. Do not write the labels, claims, and topology from reference images into a new method figure unless the user explicitly requests a faithful redraw.
Stop
Stop once the analysis results are delivered. When the user wants a prompt or an image, hand off to fig-draw. Do not compile prompts in this skill, and do not generate images.