Figure Table Planner

Figure Table Planner

Figures and tables are the evidence architecture of a paper. A good paper does not ask readers to infer the argument from a pile of plots; it shows the minimum set of artifacts needed to believe the claims.

Use this when

• The user has results and needs to decide how to present them.
• A paper draft has too many tables, redundant figures, or unclear captions.
• The main claim lacks a clear visual/table artifact.
• The user needs to split main-paper vs appendix evidence.
• The user wants to plan ablations, diagnostics, failure examples, or scaling plots.

Do not use this when

• The user needs to decide which experiments to run. Use experiment-design or hypothesis-and-ablation-planner.
• The user needs prose for the Results section. Use results-writeup after the evidence layout is clear.
• The user needs plotting code implementation details unless they explicitly ask for code.

Workflow

1. List the claims first

Before planning visuals, extract the paper claims:

• main empirical claim
• mechanism/ablation claim
• robustness or generality claim
• efficiency or cost claim
• qualitative/failure-mode claim
• theory or conceptual claim

No figure or table should exist without a claim it supports.

2. Choose the right artifact type

Use the artifact that makes the comparison easiest to judge:

• Table — exact comparisons across methods, datasets, metrics, or ablations.
• Line plot — trends over scale, time, data size, compute, or hyperparameters.
• Bar plot — small categorical comparisons when exact numbers are less important.
• Scatter plot — relationship between two continuous variables.
• Diagram — method structure, pipeline, or conceptual contribution.
• Example grid — qualitative outputs, failure modes, or interpretability cases.
• Appendix table — exhaustive results that support but do not carry the main story.

Prefer tables when readers need exact numbers. Prefer plots when shape or trend is the claim.

3. Design the main-paper evidence set

Aim for a compact sequence:

1. Method or problem diagram if the contribution is structurally hard to understand. 2. Main result table/figure that supports the headline claim. 3. Ablation table/figure that supports the mechanism claim. 4. Robustness or failure analysis that bounds the claim. 5. Efficiency/resource artifact only if efficiency is part of the contribution.

If two artifacts support the same claim in the same way, merge them or move one to appendix.

4. Specify caption jobs

A caption should tell the reader what to look for and what the artifact proves.

Caption checklist:

• names the setup and metric
• states the takeaway
• identifies higher/lower is better where relevant
• defines abbreviations
• mentions uncertainty or number of runs if relevant
• does not claim more than the artifact shows

5. Check statistical and visual hygiene

Look for:

• missing baselines
• inconsistent metrics
• unclear axes or units
• unreadable labels
• absent confidence intervals/standard deviations when variance matters
• cherry-picked examples
• color-only distinctions without labels
• appendix artifacts referenced nowhere in the main text

Output format

# Figure/Table Plan

## Claim-to-artifact map
| Claim | Artifact | Type | Main/appendix | Status |
|---|---|---|---|---|

## Recommended main-paper sequence

## Appendix evidence

## Caption drafts

## Redundant or missing artifacts

## Visual/statistical risks

Quality bar

A reader should be able to skim the figures, tables, and captions and recover the paper's argument without reading every paragraph.