Paper Figures Skill
Figures are information. A survey/review report covering downloaded papers MUST include key figures — architecture diagrams, experimental results, comparisons. Do NOT write a text-only survey when figures are available.
3-step workflow
Step 1 — Extract figures from downloaded papers
Your brain prompt supplies the script's absolute path as {{EXTRACT_FIGURES}}. Invoke it once per paper:
bash{{EXTRACT_FIGURES}} data/papers/<paper-id>.pdf # single PDF {{EXTRACT_FIGURES}} data/papers/<arxiv-id> # arXiv source directory
This creates data/papers/<id>_figures/ with extracted images and a manifest.json listing each figure with caption and page number.
Step 2 — Review captions and classify every figure
Read every manifest:
bashcat data/papers/*_figures/manifest.json
Record your decisions in notes/memory.md under a ## Figure Review section. Every figure must end up in one of three states:
- USE — essential for understanding the topic; will be included in the report
- SKIP — irrelevant, redundant, or low quality
- UNREVIEWED — caption not yet read (only acceptable as a transient state)
Select figures that are:
- Essential for understanding the topic (architecture diagrams, system schematics)
- Key experimental results that support your narrative
- Useful comparisons across methods, systems, or time periods
- Visually informative (not just tables rendered as images)
Step 3 — Include USE figures in the report
For each figure marked USE, embed it directly in LaTeX:
latex\begin{figure}[t] \centering \includegraphics[width=\linewidth]{../data/papers/<id>_figures/<filename>} \caption{<Your caption describing the figure in the context of your survey>. Adapted from \cite{<key>}.} \label{fig:<label>} \end{figure}
Rules
- Coverage: survey/review reports include at least 3-5
USEfigures from downloaded papers, in addition to any figures you generate yourself. - Captions: write your OWN caption in the context of the survey narrative — do not copy the original caption.
- Attribution: always cite the source paper with
\cite{}. - No skipping the review step: every extracted figure must be classified before writing the report.
- Complementary: you may also generate your own figures (see
skills/matplotlib-figures/SKILL.md) for data summaries, timelines, or cross-paper comparisons not found in existing figures.


