read-vis

name	read-vis
description	Check a paper's visual presentation for layout, figure quality, spacing, and formatting issues. Use when reviewing paper polish before submission.
author	PaperDoctor Research
license	MIT
argument-hint	paper.pdf
allowed-tools	Read, Bash(python *)

Paper Visual Check

Check a paper from the visual side.

Focus on what can be seen on the page:

• layout is reasonable or not
• figures are too small or too large
• spacing is awkward
• tables are cramped or hard to read
• captions are detached
• figures look low-quality or obviously AI-generated

Do not focus on code, experiments, or citations here.

When to Use

• Check whether the paper looks polished
• Find obvious visual problems before submission
• Review figures, tables, spacing, and page balance

Workflow

Visual Check:
- [ ] Resolve PDF input
- [ ] Reuse or render page images
- [ ] Inspect pages visually
- [ ] Save report

Resolve Input

• Input should be a PDF path.
• If prepare-paper has already run, use {paper_dir}/metadata/page/ as the page-image directory.
• If page images do not exist yet, render them with:

python tools/pdf_render.py /path/to/paper.pdf

This writes:

• {paper_dir}/metadata/page/page-001.png, page-002.png, ...
• {paper_dir}/metadata/page/manifest.json

What to Check

Look for visible issues such as:

• bad spacing
• poor page balance
• overflow or clipping
• a weak last line with only a few trailing words
• tiny figures or tables
• figures that are too large for their value
• captions far from the figure or table
• crowded tables
• blurry or low-quality figures
• figures with obvious AI-generation artifacts

Every finding should point to a specific target when possible:

• Figure 2
• Table 3
• right-column equation block
• bottom image on page 6
• caption under Figure 4

Do not write vague findings like "layout is a bit off". Say exactly what is wrong and where it is.

Do not treat anonymous submission formatting as a visual issue. Examples:

• Anonymous Author(s)
• hidden affiliations
• placeholder contact fields used for blind review

These are submission-state choices, not layout defects.

If a paragraph ends with a very short final line, treat that as a valid visual issue when it makes the page look poorly balanced. Suggest reflowing nearby text, adjusting spacing, or reordering content.

Zoom in when in doubt

When a figure or table is too dense or too small to judge confidently from the full page, crop the region and re-read the crop:

python -c "
from PIL import Image
Image.open('{paper_dir}/metadata/page/page-004.png').crop((LEFT, TOP, RIGHT, BOTTOM)).save('/tmp/zoom.png')
"

Then Read /tmp/zoom.png to inspect axis labels, sub-panel text, or AI-generation artifacts at full resolution. Crop instead of hedging — replace "might be too small to read" or "looks blurry but unsure" with a decisive judgment after a closer look.

Output

Write the report by issue, not by page.

Each issue should be one concrete visual problem:

{
  "page": 4,
  "quote": ["Figure 3"],
  "status": "warning",
  "reason": "Figure 3 is too small to read comfortably, especially the axis labels.",
  "suggest": "Enlarge Figure 3 or simplify the panel so the labels remain legible."
}

Use:

• page: page number
• quote: array of affected targets (e.g. ["Figure 3"] or ["Figure 4", "Figure 5"])
• status: warning or error
• reason: exact visual problem
• suggest: concrete suggestion

status rules:

• error — 100% certain defects:

  • Content clipped/cut off, text completely illegible, broken layout
  • Missing figures, factual errors in figure content (wrong labels, copy-paste)
  • Obviously AI-generated figures with visible artifacts (extra fingers, garbled text, impossible geometry, hallucinated details) used as actual scientific content (not as decorative illustrations)

• warning — genuine quality issues:

  • Rough/unpolished figures that look hastily made (misaligned elements, inconsistent styling, low-resolution rasterized text)
  • Overlapping text that obscures data, color choices that make categories indistinguishable
  • Figures or tables that are too sparse — excessive whitespace, a tiny chart floating in a large empty area, a table with 2 rows taking half a page
  • Important labels hard to read even when zoomed in

Do NOT flag (not even warning):

• Polished dense figures with clean vector graphics, consistent styling, clear legends — density is a feature when well-executed
• Tables with small but readable font — standard in conference papers
• Pages with multiple well-formatted figures/tables — layout choice

Quality judgment guide:

• Polished & informative (clean vector graphics, consistent colors, clear legends, professional layout) → do not flag, even if dense
• Rough & poorly executed (default matplotlib with no styling, blurry/pixelated sub-images, misaligned labels, hard to compare panels, missing visual separators, low-resolution rasterized content, inconsistent fonts/sizing) → warning
• AI-generated artifacts (DALL-E/Midjourney figures with telltale distortions used as scientific figures) → error
• Too sparse (a bar chart with 3 bars taking an entire page, mostly empty figure area) → warning
• Dense comparison grids where sub-images are too small to see meaningful differences → warning

Font size rule of thumb: Compare text inside figures/tables to the paper's body text. If figure labels, axis text, or table content is noticeably smaller than the body text (roughly <60% of body font size), flag as warning. Body text in a standard conference paper is ~10pt; figure text below ~6pt is a problem.

Figure-caption coherence:

• A good figure makes its caption easy to understand — the visual clearly illustrates what the caption describes
• warning if a figure is hard to connect to its caption (e.g. caption describes a trend but the figure doesn't clearly show it, or caption references sub-panels by labels that are missing/hard to find)
• error if the caption describes something completely different from what the figure shows (factual mismatch)

Aim for quality over quantity — fewer, more actionable findings are better than many nitpicks.

Save one report with:

• summary
• issues

Example shape:

{
  "summary": {
    "total_issues": 5,
    "warning_issues": 4,
    "error_issues": 1
  },
  "results": [
    {
      "page": 4,
      "quote": ["Figure 3"],
      "status": "warning",
      "reason": "Figure 3 is too small to read comfortably, especially the axis labels.",
      "suggest": "Enlarge Figure 3 or simplify the panel so the labels remain legible."
    }
  ]
}

If a page has no visual problem, do not create a dummy entry for that page.

Output:

• {paper_dir}/reports/check_vis.json

Tips

• Judge from visual evidence only.
• Prefer obvious, actionable feedback.
• If a figure looks AI-generated, say what visual artifact suggests that.