Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when writing or fixing LaTeX manuscript source: apply academic typography conventions (en-dashes, TeX quotes, booktabs, ~ before refs) the default would not.
.claude/skills/latex-typography-conventions/SKILL.md| Model | Eval pass | Runs |
|---|---|---|
| gemini-3.6-flash | 100% | 4 |
| Model | Lift | Δ tokens | Δ turns | Cases | Verified |
|---|---|---|---|---|---|
| gemini-3.6-flashbest | +32% | +362% | 0% | 22 | 54d ago |
| gemini-3.5-flash | pending re-run | — | |||
| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-15 | ✗→✓ | ▲ Improved | — | — |
| case-01 | ✗→✓ | ▲ Improved | — | — |
| case-06 | ✗→✓ | ▲ Improved | — | — |
| case-02 | ✗→✓ | ▲ Improved | — | — |
| case-04 | ✗→✓ | ▲ Improved | — | — |
When you emit LaTeX source for an academic manuscript, produce it in the cross-venue typographic conventions below. Apply this any time you write or fix .tex content — tables, figures, prose, math, references, units. These are exact, arbitrary TeX-specific forms; the naive default (straight quotes, \hline, 10-20, ..., $$) is what an inexperienced author types and what reviewers read as amateur typesetting.
booktabs rules \toprule (top), \midrule (under the header),\bottomrule (bottom). Never \hline.
\hline\hline, or \toprule\toprule).\midrule separates the header row from the body; do not rule betweenevery data row.
tabular/array column spec carries letters only: {lcr}, {llc}.| in the spec ({|l|c|r|}). Vertical rules in atable are the single clearest amateur tell.
r) or decimal-aligned (S fromsiunitx); text columns are left-aligned (l).
\caption{} BEFORE \begin{tabular}, inside the table float.\caption{} AFTER \includegraphics, inside the figure float.[htbp][htbp] (or [tbp], [t]) on \begin{table} / \begin{figure}.[H] (forces placement, creates large whitespace gaps) and never[h!]. Plain [h] alone is fragile — avoid it too.
--10--20,pp.\ 5--12, 2020--2023. Never a single hyphen 10-20 for a range.
--- for asides, hyphen for compoundstext---text. Apply spacing uniformly (closed text---text is the academic default).
- is only for compound words (state-of-the-art,well-known). Never use a single hyphen where a range (R6) or aside belongs.
and close with two apostrophes''. Never the straight " keyboard character — it renders as two wrong glyphs in TeX.
He called it optimal' which is debatable.''
\ldots\ldots (or \dots). Never three literal periods ...,which TeX sets with wrong inter-dot spacing.
~ immediately before every cross-reference andcitation: Figure~\ref{fig:x}, Table~\ref{tab:x}, Section~\ref{sec:x}, Equation~\eqref{eq:x}, and prior work~\cite{smith2020}.
end of one line and the number on the next.
\,: 10\,ms, 8\,GB,100\,K. Equivalently use siunitx: \SI{10}{\milli\second} / \qty{10}{ms}.
10ms (no space) and never 10 ms (a plain inter-word space, whichTeX may break across lines).
\nicefrac\nicefrac{1}{2} (or\sfrac{1}{2} from xfrac). Never an inline $\frac{1}{2}$, which creates a tall element that disrupts line spacing.
\frac is correct only inside display math (R13).\[ ... \]\[ ... \] or an equation/align environment.$$ ... $$ — that is plain TeX, not LaTeX, and produces incorrectvertical spacing.
$\times$ (e.g. a $3 \times 3$ kernel). Never the letterx.
$^\circ$ or \textdegree (or a siunitx celsiuscommand). Never the Unicode ° character.
\emph{}\emph{...}. Never \textbf, \textit, ALL CAPS, or acolor command for stress in running prose — \emph is semantic and nests correctly.
\&, \%, \#, \_ whenever they appear as literal text. A bare&, %, #, or _ is a TeX control character and breaks compilation or is silently dropped.
\url{}\url{...} or \href{...}{...}. Never leave it as bare textand never put it in \texttt{...} — only \url allows the line breaks that prevent an overfull margin.
R1/R2/R3 — table:
latex% BEFORE \begin{table}[H] \begin{tabular}{|l|c|r|} \hline Method & Accuracy & F1 \\ \hline\hline Ours & 0.94 & 0.91 \\ \hline \end{tabular} \caption{Results.} \end{table} % AFTER \begin{table}[htbp] \caption{Results.} \begin{tabular}{lcr} \toprule Method & Accuracy & F1 \\ \midrule Ours & 0.94 & 0.91 \\ \bottomrule \end{tabular} \end{table}
R4/R5 — figure:
latex% BEFORE \begin{figure}[H] \caption{Training curves over 100 epochs} \includegraphics{results.png} \end{figure} % AFTER \begin{figure}[htbp] \includegraphics{results.png} \caption{Training curves over 100 epochs} \end{figure}
R6 — number range: BEFORE See pages 10-20. → AFTER See pages 10--20.
R6 — year range: BEFORE the years 2020-2023 → AFTER the years 2020--2023
R7 — aside: BEFORE works well, in fact better than expected, on all → AFTER works well---in fact better than expected---on all
R8 — quotes: BEFORE the result was "surprising" → AFTER the result was surprising''
R8 — nested: BEFORE She said "He called it 'optimal'." → AFTER She said He called it optimal'.''
R9 — ellipsis: BEFORE the analysis trailed off... → AFTER the analysis trailed off\ldots
R10 — ref: BEFORE As shown in Figure \ref{fig:curve} → AFTER As shown in Figure~\ref{fig:curve}
R10 — cite: BEFORE Prior work \cite{smith2020} → AFTER Prior work~\cite{smith2020}
R11 — unit: BEFORE latency was 10 ms or 10ms → AFTER latency was 10\,ms
R11 — unit: BEFORE under 8 GB → AFTER under 8\,GB
R12 — fraction: BEFORE about $\frac{1}{2}$ of the runs → AFTER about \nicefrac{1}{2} of the runs
R13 — display math: BEFORE $$L = \sum_i (y_i - \hat{y}_i)^2$$ → AFTER \[ L = \sum_i (y_i - \hat{y}_i)^2 \]
R14 — multiply: BEFORE a 3 x 3 convolution → AFTER a $3 \times 3$ convolution
R14 — degree: BEFORE at 25°C → AFTER at $25^\circ$C
R15 — emphasis: BEFORE This step is \textbf{critical}. → AFTER This step is \emph{critical}.
R16 — escapes: BEFORE Smith & Jones, a 5% gain → AFTER Smith \& Jones, a 5\% gain
R16 — underscore: BEFORE the field max_value → AFTER the field max\_value
R17 — URL: BEFORE available at https://example.com/repo/path → AFTER available at \url{https://example.com/repo/path}
- joins acompound word, en-dash -- spans a range, em-dash --- breaks a clause. When in doubt about a numeric "A to B," it is a range → --.
(e.g. IEEEtran), follow it rather than imposing booktabs — but the default case is always booktabs.
% after a number conventionally touches it(14.3\%); a unit like ms/GB does not (10\,ms). Either way % is still escaped as \%.
\frac is right in display math. R12 forbids inline $\frac{}{}$, not\frac itself — inside \[ ... \] or equation, \frac is correct.
et al., vs., e.g., i.e. are set upright (notitalic) and take a period; this is orthogonal to R15 emphasis.
mode ($-5$), not a text hyphen.
\hline; always \toprule/\midrule/\bottomrule.| in a column spec; always letters only ({lcr}).[H] or [h!]; always [htbp].--."; always ...'' ....; always \ldots.Figure \ref{} with a space; always Figure~\ref{} with a tilde.10 ms or 10ms; always 10\,ms.$\frac{}{}$; always \nicefrac{}{}.$$...$$; always \[...\].x for multiply or ° for degree; always $\times$ and $^\circ$.\textbf/\textit/color for prose emphasis; always \emph{}.&, %, #, _; always \&, \%, \#, \_.\texttt{} URL; always \url{}.\hline and | columns — the universal novice default.10-20 for ranges; it reads as a subtraction."..." and literal ... out of word-processor habit.\ref/\cite — the most common LaTeX mistake.10ms) or using a breakable plain space.$\frac{1}{2}$ mid-sentence and $$...$$ for display math.\textbf{} for emphasis because it "looks emphasized."&/%/#/_, which silently breaks the build.booktabs rules · no | columns · table caption above · figure caption below · [htbp] · en-dash ranges · em-dash asides · ...'' quotes · \ldots · tilde before refs/cites · 10\,ms units · \nicefrac · \[...\] · $\times$ / $^\circ$ · \emph{} · escaped \& \% \# \_ · \url{}.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-15 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-01 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-12 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-06 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-02 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-04 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-11 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-17 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-10 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-03 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-20 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-07 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-05 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-16 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-21 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-09 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-14 | pass→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-08 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-13 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-18 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-19 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-22 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +32 percentage points is the difference between those two pass rates over the 22 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
The per-case answers from this run were removed by the retention sweep, so the case table below shows the verdicts without the text either arm produced. The counts above were recorded at the time and are unaffected. Answers are now kept for 180 days.
| Model | Method | Date | Lift |
|---|---|---|---|
| gemini-3.5-flash | verified | 7/9/2026 | +21% |
Other measured skills in the registry, with their headline benchmark lift.