figure-table-audit - SKILL.md Agent Skill

name: figure-table-audit description: Audit figures, tables, captions, cross-references, and statistical notes. argument-hint: "[path to manuscript, figures, tables, SI, or compiled PDF; include target journal if known]"

Figure and Table Auditor

Heritage and scope

This is an original Open Science Skills workflow for manuscript QA. It remixes general figure/table and citation-compliance ideas from Cheng-I Wu's Academic Research Skills for Claude Code (CC BY-NC 4.0), but is rewritten for open-science social-science manuscripts. It is not a visual hallucination engine: when a claim requires reading plotted values from an image, prefer source data or mark the issue as needing author verification.

This is the end-stage auditor. For figure design and production guidance during drafting, use the figures skill; for table design, use the tables skill. Run figure-table-audit once the figure and table set is stable and you are preparing for submission.

Instructions

1. Orient before auditing

Identify:

Manuscript source and compiled output, if any.
Figure directory, table files, appendix/SI files, and build command.
Target journal or style guide.
Whether figures are generated from code, manually edited, or exported from software.
Whether source data for figures/tables are available.

If only a PDF is available, state that cross-reference and value checks are lower confidence.

2. Inventory figures and tables

Build an inventory with:

Figure/table number or label.
File path or source location.
Caption/title.
First in-text callout.
Appendix/SI location if applicable.
Producing script or data source, if visible.

Check:

Every in-text callout resolves to an existing figure/table.
Every figure/table is called out in text before or near appearance.
Numbering is gap-free and not duplicated.
Main-text and SI labels do not collide.
LaTeX/Markdown references compile without unresolved labels.

3. Check text-to-evidence consistency

For each figure/table used to support a substantive claim:

Match the in-text claim to the exact row, column, panel, model, or plotted quantity.
Verify direction, magnitude, uncertainty, subgroup, and denominator.
Check whether text overstates non-significant or imprecise estimates.
Check whether figure/table notes disclose model specification, N, weights, fixed effects, clustering, and missing-data handling.
For experimental papers, check ITT vs per-protocol labeling, attrition-by-arm, baseline balance, and CONSORT/sample-flow consistency.

Do not infer exact values by eyeballing a plot unless the figure encodes labeled values. If source data are unavailable, write VISUAL READ ONLY - AUTHOR VERIFY.

4. Audit captions and notes

Captions and table notes should let a reader understand the evidence without hunting:

Figure caption states what is plotted, units, sample, uncertainty interval, and panel meanings.
Table title states the estimand or model family, not only "Results."
Table notes define dependent variable, treatment/condition coding, omitted categories, controls, fixed effects, clustering, weights, and significance markers.
All abbreviations are defined on first use.
All transformations, scales, and index directions are explicit.
Any data exclusions or subgroup restrictions are named.

For conjoint, list-experiment, topic-modeling, LLM-classification, and OCR studies, invoke or recommend the relevant sibling skill when table/figure interpretation depends on method-specific standards.

5. Check accessibility and production quality

Flag:

Color palettes not interpretable in grayscale or by color-blind readers.
Low-resolution or rasterized text in vector outputs.
Axis labels too small or missing units.
Legends that obscure data or use ambiguous labels.
Inconsistent decimal precision.
Overcrowded tables that should move to SI.
Missing alt text when the target venue requires it.
Non-reproducible manual edits that are not documented.

6. Audit SI and replication linkage

Check whether:

Main-text claims that point to SI land on the correct appendix/table/figure.
SI numbering and captions are internally consistent.
Each main figure/table can be traced to a script or documented manual step.
Replication README maps outputs to scripts and data files.
Figure/table source data are archived or clearly restricted.

Output

Produce a Figure and Table Audit Report:

# Figure and Table Audit Report

Scope:
Inputs checked:
Build/source status:
Summary: <N blocking, N recommended, N minor, N author-verification>

## Inventory
| ID | Path/location | Caption/title | First callout | Source/script |

## Blocking Issues
| Location | Figure/table | Issue | Evidence | Fix |

## Recommended Fixes
| Location | Figure/table | Issue | Fix |

## Minor / Production Issues
| Figure/table | Issue | Fix |

## Author Verification Needed
| Figure/table | Why verification is needed |

## Readiness Checklist
| Dimension | PASS/FAIL/PARTIAL/NA | Notes |

Severity:

Blocking: missing figure/table, wrong referenced value, denominator mismatch, unresolved label, table contradicts text, missing sample-flow evidence for experimental paper, or non-reproducible main result output.
Recommended: incomplete notes, unclear captions, missing units, missing source script, accessibility problem, imprecise uncertainty reporting.
Minor: style, spacing, decimal precision, typography, cosmetic consistency.

Quality checks

Figure/table inventory was built before findings were listed.
Every substantive text-to-table finding names the row/column/panel/model checked.
Visual-only readings are flagged for author verification unless exact values are labeled.
Captions and notes were checked for sample, units, uncertainty, and model details.
SI and replication links were checked when files were available.
Method-specific figures/tables triggered the relevant sibling skill when needed.