Figure Caption Reader
Purpose
Convert paper figures, screenshots, plots, and captions into technical understanding. The output should explain what the visual evidence is trying to prove and what a reader should verify.
When To Use
Use this skill when the task includes a paper figure, chart, table screenshot, qualitative comparison, pipeline diagram, attention map, architecture figure, or result visualization.
Inputs
- Image, screenshot, rendered PDF page, or figure crop.
- Caption or surrounding paper text if available.
- Optional paper title and task.
- Optional local experiment logs or baseline numbers.
Workflow
- Identify the visual type: architecture, qualitative result, quantitative plot, ablation table, failure case, dataset example, or workflow diagram.
- Describe only what is visible first. Avoid inventing labels or values that are not legible.
- Extract the intended claim: what the authors want the figure to prove.
- Link visual elements to technical concepts: modules, loss terms, data paths, prompts, memory, policy, or benchmark splits.
- For plots and tables, state the axes, trends, outliers, and comparison groups.
- For qualitative figures, separate visual impression from measurable evidence.
- End with a verification checklist: what code, metric, or extra table would confirm the figure's claim.
Output
Visible ContentIntended ClaimTechnical ReadingEvidence StrengthVerification Checklist
Quality Gates
- Never hallucinate unreadable numbers.
- If text is too small, say so and request a higher-resolution crop or PDF page.
- Distinguish visual quality from metric superiority.
- Do not treat a single cherry-picked qualitative image as proof of general performance.
微信扫一扫