Shiro Kuriwaki’s instructions to LLMs for LaTeX tables

Use these defaults for descriptive and regression tables in social science papers and results memos. Explicit user requests and journal requirements take precedence. Keep analytical choices separate from presentation preferences.

Captions and titles

Number tables automatically with \caption, not \caption*. Use a colon: Table 1: Sample Construction. Make both the label and short title bold and normal text size, even when the table body uses smaller text.

Use concise, substantive titles and consistent wording across the paper.

Put sample definitions, estimator details, and qualifications in normal-weight notes. If explanatory prose follows the short title inside a caption, reset that prose to \normalfont. Do not repeat a title inside the table body.

Keep captions, labels, and notes in one maintained source. Follow the existing division between the R generator and manuscript wrapper; avoid duplicate captions or manual edits to generated tables.

When a memo’s AI acknowledgment is requested as a header, use the requested wording in a running header on every page. Do not add it to papers by default.

\usepackage{booktabs,caption,array,makecell}
\captionsetup[table]{
  font={normalsize,bf},labelfont=bf,textfont=bf,labelsep=colon,skip=6pt
}

Width, alignment, and page layout

Use the table’s natural width when it fits. Do not shrink a short table to \linewidth or add whitespace merely to fill the page. Start with tabular rather than a width-filling tabular*.

Wrap long or awkward column headings deliberately, especially when headings create wide columns with mostly empty bodies. Use short, interpretable labels such as “Has newsletter in window”; \makecell{Has newsletter\\in window} is one way to wrap them.

Left-align row labels and prose. In regression tables, center coefficients, standard errors, sample sizes, and other numeric or categorical model entries, including Yes/No indicators. Descriptive counts may be right-aligned.

Use booktabs rules to separate headers and meaningful panels. Avoid vertical rules, boxed grids, and decorative shading as defaults.

Let tables flow naturally. Do not insert a page break after every table. Keep a regression table and its notes together when feasible.

For many paired outcomes, start with four outcomes per panel, each with two adjacent model columns. Stack panels on one page; use landscape and a modest reduction to \small if helpful. Adapt to the actual number of outcomes. Avoid scaling the entire table to illegibly small text with \resizebox.

Use normal body text size when practical. Smaller table text does not imply smaller captions. Keep notes readable and spacing modest.

Numbers, labels, and sample definitions

Display rates as proportions with a leading zero and three decimals: 0.215, not 21.5% or .22. Use thousands separators for counts: 21,437.

Keep regression coefficients and SEs in their estimation units, usually probability units for linear probability models. Do not multiply them by 100 unless percentage points are explicitly requested and labeled.

Default coefficients and SEs to three decimals; increase precision when needed to preserve meaningful differences. Explain parentheses and any significance stars in the notes, including thresholds and the clustering unit.

Use a Unicode check mark ✓ for included fixed effects and a blank cell for excluded fixed effects. Ensure the TeX engine/font renders the glyph. A blank FE cell means exclusion, not missing data; use a dash for genuinely missing statistics.

State the numerator and denominator of rates; do not silently code missing as zero.

Fold brief row-specific qualifications into the relevant cell when that eliminates a redundant column: 70,228 (hypothetical) or 21,437 (actual).

Model comparisons

Group related models under the outcome name. Use short column labels and explain model distinctions once in the notes.

Report sample size as “N” in every regression column. For regressions with clustered standard errors, also report the number of clusters and identify the clustering unit in the notes.

Include outcome means and FE indicators when relevant. Omit likelihoods, scores, AIC/BIC, and similar diagnostics unless requested.

Keep variable definitions, units, and sample restrictions consistent when a comparison is meant to isolate a modeling choice. Disclose differences that remain; a figure described as a visualization of a table must use that table’s actual estimates, samples, and uncertainty measures.

Generation and verification

Use gt for descriptive tables and modelsummary::modelsummary() for regression tables. Generate both TeX and a gt HTML preview with an editable gt object, unless the user requests fewer formats.

Update the generator and regenerate outputs. Use stable filenames and project-relative TeX inputs that work in the manuscript or Overleaf.

Write R in tidyverse style, with descriptive section headings and lines no longer than 80 characters. Use glue, scales, and readr I/O; specify join relationships. New Codex-authored scripts start with # Written by Codex.

Keep fast intermediate models and summaries in memory unless a later script explicitly reads them or they are expensive to recompute. Editable gt tables are final deliverables, not disposable model caches.

Compile with the manuscript’s actual page size and fonts, then inspect the rendered PDF. HTML styling is not proof of correct TeX styling. Check caption numbering, FE glyphs, alignment, wrapping, panel fit, notes, and page breaks; resolve clipped text and overfull boxes.