In short
A famous AI paper, shortened and laid out again. Its example prompts and the model's answers sit in labelled boxes side by side, with a tick or a cross on each answer, and the charts are drawn from the paper's numbers.
What you'll build
Seven pages of a machine-learning preprint on US letter: Wei et al.'s Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, abridged from its arXiv version under CC BY 4.0. The first page carries the paper's best-known figure, set as boxes rather than pasted as a picture: standard prompting on the left, chain-of-thought prompting on the right, each prompt in a box tabbed Model Input and each answer in a box tabbed Model Output, the exemplars in IBM Plex Mono and the chain of thought in bold colour, with a cross on the wrong answer and a tick on the right one. The text cites the way NeurIPS papers do, (Brown et al., 2020), from the paper's own reference list rebuilt as BibTeX. Table 2 and two charts are redrawn from the numbers the paper publishes.
This recipe answers
- How do I set an AI paper's prompt and model-output examples in boxes side by side?
- How do I re-set an open-access paper from arXiv or PubMed Central with Postext, keeping its citations, figures and licence line?
- How do I cite works and build the bibliography in APA, IEEE or another citation style?
The short answer
// :::columns{count=2 breaks="4"} inside the "figure" box opens the right column at its fourth
// block; each nested box counts as one block.
const tab = (fill) => ({ fontFamily: SANS, fontSize: pt(7), fontWeight: 600,
color: col('paper'), background: col(fill), position: 'top-left', inset: mm(3),
height: mm(4.2), offset: mm(2.1), paddingX: mm(2.2) }); // straddles the top edge
const exemplar = (id, fill, ink, extra) => ({ id, background: col(fill),
border: { enabled: true, color: col('rule'), width: pt(0.6) }, borderRadius: mm(2.4),
padding: { top: mm(4.4), right: mm(3), bottom: mm(2.6), left: mm(3) },
marginTop: mm(4.2), marginBottom: ZERO, label: tab(ink),
body: { fontFamily: MONO, fontSize: pt(7.8), lineHeight: pt(10.4), color: col('ink'),
boldColor: col(ink), // **…** marks the chain of thought: the highlight of the original
textAlign: 'left', hyphenation: false, paragraphSpacing: true, firstLineIndent: ZERO },
...extra });
const mark = (id) => ({ icon: { kind: 'resource', resourceId: id, size: mm(5.6),
position: 'corner', cornerSide: 'right' } }); // a badge on the top-right corner
const promptBoxes = [
{ id: 'figure', span: 'page', backgroundEnabled: false, border: { enabled: false },
padding: mm(0), columnGap: mm(6), marginTop: ZERO, marginBottom: pt(LEAD),
body: { fontFamily: SANS, fontSize: pt(8.4), lineHeight: pt(11.4), color: col('ink'),
boldColor: col('accent'), textAlign: 'left', paragraphSpacing: true,
firstLineIndent: ZERO } },
exemplar('input', 'tint', 'accent'),
exemplar('right', 'mint', 'green', mark('mark-right')),
exemplar('wrong', 'mint', 'green', mark('mark-wrong')),
];
Figure 1 as boxes: two columns, a tab on each box, a ✓ or ✗ in the corner
Ingredients
- Features
- Columns in a box or in the textNested boxesNumbered box tabsBox icons and corner badgesCallout boxesCitations in a citation styleBibliography from the referencesCross-referencesNumbered headingsHeading stylesDesigned openersCustom resource typesCaption styleTables from dataTable styleFootnotesLink addresses in printRunning heads and foliosSemantic colour palettePDF export
- Also uses
- Citations that place figuresHeading attributesHeads by page roleParagraph stylesFonts embedded in the PDFFigures and tables as resourcesLine breaks in titlesUnnumbered chapters
- Type
- Newsreader, IBM Plex Sans, IBM Plex Mono (SIL OFL 1.1)
- Assets
- None: every picture is drawn in code
Method
#1 · Figure 1 is three box styles
The code is the short answer above. The figure is one page-wide box with no frame of its own; inside it, :::columns{count=2 breaks="4"} opens the right column at the fourth block, and every nested box counts as one block. The prompt and answer boxes share one style factory, so they differ only in fill, tab colour and the badge on the corner: a callout icon with position: 'corner' takes an SVG resource, which is how the ✓ and ✗ reach the page in a document whose fonts have no such glyphs. In your own paper, write each exemplar as plain paragraphs inside a box and mark the reasoning with **…**; the box style's boldColor turns that into the highlight.
#2 · Author–year citations, natbib style
registerCitationEngine(createCiteprocEngine({ styles: STYLES, locales: LOCALES }));
// Chicago author-date writes (Brown et al. 2020); ML venues put a comma before the year.
const AUTHOR_YEAR = '<group delimiter=" ">\n <text macro="author-inline"/>';
const natbib = STYLES['chicago-author-date']
.replace(AUTHOR_YEAR, AUTHOR_YEAR.replace('" "', '", "'));
const citations = { style: 'custom', customStyle: natbib, link: true,
bibliography: { fontSize: em(0.8), lineHeight: pt(10.2), entrySpacing: pt(1.8),
hangingIndent: mm(4), doi: 'text' } };
// The text's two web addresses are links: clickable in the PDF, printed in parentheses on
// paper, `[name census data](https://namecensus.com/)` as "name census data (namecensus.com)".
const links = { print: 'inline', urlDisplay: 'short' };
ML venues use natbib's author–year look: a comma between the names and the year, et al. from three authors on. Chicago author-date is the bundled style closest to it, so the recipe edits one group of its CSL, the one that joins author and year, from a space to a comma. Bring your own .bib file as it is: paste it into a :::references{format=bibtex} block, cite with [@key] for (Brown et al., 2020) and @key for Brown et al. (2020). A suffix keeps the comma written before it and its emphasis, so [@brown2020language, *inter alia*] prints (Brown et al., 2020, inter alia) as the paper does. Captions cite the same way: Figure 4's caption writes @jie2022learning and @lan2021mwptoolkit, and both works enter the reference list with no nocite.
The two web addresses in the text, the name census data in section 5 and the licence in the colophon, are written as links, [name census data](https://namecensus.com/). With links: { print: 'inline', urlDisplay: 'short' } the paper prints each address in parentheses after its text, without https://, and the PDF keeps the link. An address typed as plain text prints the same but goes nowhere when clicked.
#3 · The title block
const text = (id, content, family, size, extra) => ({ kind: 'text', id, content, align: 'left',
fontFamily: family, fontSize: pt(size), color: col('ink'), overflow: 'wrap', ...extra });
const at = (to, edge, y, width = MEASURE) => ({ anchor: { to, edge },
offset: { x: ZERO, y: mm(y) }, size: { width: mm(width), height: 'auto' } });
const titleBlock = { enabled: true, minHeight: mm(58), slot: { elements: [
text('venue', '{attr.venue}', MONO, 7.5, { fontWeight: 500, letterSpacing: pt(0.6),
color: col('accent'), placement: at('container', 'top-left', 0) }),
text('title', '{titleText}', SERIF, 23, { fontWeight: 600, lineHeight: 1.1,
placement: at('#venue', 'below', 5) }),
text('authors', '{attr.authors}', SANS, 9.6, { fontWeight: 500, lineHeight: 1.45,
placement: at('#title', 'below', 6) }),
text('affiliation', '{attr.affiliation}', SANS, 8.6, { color: col('muted'),
placement: at('#authors', 'below', 1.2) }),
{ kind: 'rule', id: 'rule', thickness: pt(0.5), color: col('rule'),
placement: at('#affiliation', 'below', 4) },
text('note', '{attr.note}', SERIF, 8.2, { lineHeight: 1.35,
color: col('muted'), placement: at('#rule', 'below', 2) }),
] } };
The title is the document's only level-1 heading, with a style that spans both columns and draws a designed block: venue line, title, authors in two rows, affiliation and a note on the abridgement. The authors and the note are heading attributes, and \n inside an attribute starts a new row, as \\ does in a LaTeX \author. For your own paper, change the attributes; the design stays.
#4 · Figure numbers that match the original
// Figure 1 is boxes, which the figure counter does not see: number the charts by type.
const numbered = (id, word, n, extra) => ({ id, name: `${word} ${n}`,
shortLabel: `${word} ${n}`, captionPrefix: `${word} ${n}`, numberingTemplate: '',
resetOn: 'never', counterFormat: 'decimal', ...extra });
const resourceTypes = [numbered('mark', 'Mark', ''), numbered('fig2', 'Figure', 2),
numbered('fig4', 'Figure', 4),
numbered('tab2', 'Table', 2, { captionStyle: { position: 'above' } })];
const ACROSS = { position: 'top', span: 'page' }; // a float over both columns
const svg = (id, typeId, file, [width, height], caption, note, altText, placement) => ({
id, typeId, kind: 'svg', createdAt: 0, updatedAt: 0,
placement: placement ?? { position: 'top' }, svg: { fileId: file, width, height },
caption, note, altText });
The engine numbers figures in the order they are first cited, but Figure 1 here is a box, which the counter never sees. So the redrawn charts get a resource type each, with the number written into the label and an empty numberingTemplate, and keep the numbers of the arXiv version: Figure 2, Figure 4 and Table 2. In a paper of your own, where every figure is a resource, one figure type with numberingTemplate: '{n}' does the counting.
#5 · Table 2 from the paper's data
const SETS = ['GSM8K', 'SVAMP', 'ASDiv', 'AQuA', 'MAWPS'];
const RESULTS = [ // model, size, then standard / CoT pairs per benchmark; * CoT beats standard
['UL2', '20B', '4.1 4.4* 10.1 12.5* 16.0 16.9* 20.5 23.6* 16.6 19.1*'],
['LaMDA', '420M', '2.6 0.4 2.5 1.6 3.2 0.8 23.5 8.3 3.2 0.9'],
['', '2B', '3.6 1.9 3.3 2.4 4.1 3.8 22.9 17.7 3.9 3.1'],
['', '8B', '3.2 1.6 4.3 3.4 5.9 5.0 22.8 18.6 5.3 4.8'],
['', '68B', '5.7 8.2* 13.6 18.8* 21.8 23.1* 22.3 20.2 21.6 30.6*'],
['', '137B', '6.5 14.3* 29.5 37.5* 40.1 46.6* 25.5 20.6 43.2 57.9*'],
['GPT', '350M', '2.2 0.5 1.4 0.8 2.1 0.8 18.1 8.7 2.4 1.1'],
['', '1.3B', '2.4 0.5 1.5 1.7 2.6 1.4 12.6 4.3 3.1 1.7'],
['', '6.7B', '4.0 2.4 6.1 3.1 8.6 3.6 15.4 13.4 8.8 3.5'],
['', '175B', '15.6 46.9* 65.7 68.9* 70.3 71.3* 24.8 35.8* 72.7 87.1*'],
['Codex', '–', '19.7 63.1* 69.9 76.4* 74.0 80.4* 29.5 45.3* 78.7 92.6*'],
['PaLM', '8B', '4.9 4.1 15.1 16.8* 23.7 25.2* 19.3 21.7* 26.2 30.5*'],
['', '62B', '9.6 29.9* 48.2 46.7 58.7 61.9* 25.6 22.4 61.8 80.3*'],
['', '540B', '17.9 56.9* 69.4 79.0* 72.1 73.9* 25.2 35.8* 79.2 93.3*'],
];
const cell = (content, extra) => ({ content, align: 'right', ...extra });
const tableRows = () => [
[cell('Model', { isHeader: true, align: 'left', colSpan: 2 }), { content: '',
hiddenBy: { row: 0, col: 0 } }, // gotcha: merged-cells-hiddenby
...SETS.flatMap((set, i) => [cell(set, { isHeader: true, align: 'center', colSpan: 2 }),
{ content: '', hiddenBy: { row: 0, col: 2 + 2 * i } }])],
[cell('', { isHeader: true }), cell('', { isHeader: true }),
...SETS.flatMap(() => [cell('standard', { isHeader: true }), cell('CoT', { isHeader: true })])],
...RESULTS.map(([model, size, values]) => [cell(model ? `**${model}**` : '', { align: 'left' }),
cell(size, { align: 'left' }),
...values.split(' ').map((v) => cell(v.endsWith('*') ? `**${v.slice(0, -1)}**` : v))]),
];
The results are typed once, as the LaTeX source has them, and feed both Table 2 and Figure 4, so the chart cannot drift from the table. The benchmark names span two columns each; a merged cell needs a placeholder marked hiddenBy where it spreads. Bold marks the cells where chain-of-thought prompting beats standard prompting, as the green cells of the original do.
The whole recipe
// ═══ Postext Cookbook · Nº 137 · An AI preprint with prompt exemplars in boxes ═══════
// https://postext.dev/en/cookbook/ai-preprint-prompt-exemplars
// Code: MIT · Text: Wei et al. 2022, arXiv:2201.11903 (CC BY 4.0) · Charts: drawn in code
// Fonts: Newsreader, IBM Plex Sans, IBM Plex Mono (SIL OFL 1.1) · Needs postext ≥ 1.65.0
import {
buildDocumentWithFonts, renderPageToCanvas, registerCitationEngine, registerResourceImage,
inlineSvgFonts,
} from 'https://esm.sh/postext';
import { renderToPdf, decompressWoff2 } from 'https://esm.sh/postext-pdf';
import { createCiteprocEngine, STYLES, LOCALES } from 'https://esm.sh/postext-citeproc';
const LANG = 'en'; // @lang: the language of the sample document ('en' | 'es')
const RECIPE = 'ai-preprint-prompt-exemplars';
// ─── 1 · Design ─────────────────────────────────────────────────────────────
// #region palette: a cool ink, a prompt blue, an answer green and a red for the wrong answer
const palette = { ink: '#1b1e24', accent: '#1f5fa6', green: '#21744a', red: '#b3362d',
tint: '#e9f0f8', mint: '#e6f2eb', rule: '#b4bcc8', muted: '#5b6270', paper: '#ffffff' };
const col = (id) => ({ hex: palette[id], model: 'hex', paletteId: id });
const colorPalette = Object.entries({ ...palette, 'main-color': palette.accent })
.map(([id, hex]) => ({ id, name: id, value: { hex, model: 'hex' } }));
// #endregion
const [SERIF, SANS, MONO] = ['Newsreader', 'IBM Plex Sans', 'IBM Plex Mono'];
// US letter, two columns of 85.5 mm as ML venues print.
const [TRIM_W, TRIM_H, TOP, BOTTOM, SIDE, GUTTER] = [215.9, 279.4, 22, 22, 19, 7];
const MEASURE = TRIM_W - 2 * SIDE;
const COLUMN = (MEASURE - GUTTER) / 2;
const LEAD = 12.8; // pt
const ZERO = pt(0);
// #region answer: Figure 1 as boxes: two columns, a tab on each box, a ✓ or ✗ in the corner
// :::columns{count=2 breaks="4"} inside the "figure" box opens the right column at its fourth
// block; each nested box counts as one block.
const tab = (fill) => ({ fontFamily: SANS, fontSize: pt(7), fontWeight: 600,
color: col('paper'), background: col(fill), position: 'top-left', inset: mm(3),
height: mm(4.2), offset: mm(2.1), paddingX: mm(2.2) }); // straddles the top edge
const exemplar = (id, fill, ink, extra) => ({ id, background: col(fill),
border: { enabled: true, color: col('rule'), width: pt(0.6) }, borderRadius: mm(2.4),
padding: { top: mm(4.4), right: mm(3), bottom: mm(2.6), left: mm(3) },
marginTop: mm(4.2), marginBottom: ZERO, label: tab(ink),
body: { fontFamily: MONO, fontSize: pt(7.8), lineHeight: pt(10.4), color: col('ink'),
boldColor: col(ink), // **…** marks the chain of thought: the highlight of the original
textAlign: 'left', hyphenation: false, paragraphSpacing: true, firstLineIndent: ZERO },
...extra });
const mark = (id) => ({ icon: { kind: 'resource', resourceId: id, size: mm(5.6),
position: 'corner', cornerSide: 'right' } }); // a badge on the top-right corner
const promptBoxes = [
{ id: 'figure', span: 'page', backgroundEnabled: false, border: { enabled: false },
The composed script.js runs as it is: paste it into any page’s module script, or open the recipe on CodePen. Recipe folder on GitHub ↗ (opens in a new tab)
Variations
#Cite in APA
APA writes (Roy & Roth, 2015) and lists up to twenty authors, which the bibliography of a large-model paper rarely wants.
-const citations = { style: 'custom', customStyle: natbib, link: true,
+const citations = { style: 'apa', link: true,#Set the paper in one column
A NeurIPS-style single column needs wider side margins to keep a readable measure; the page-wide boxes then take the single column.
-const [TRIM_W, TRIM_H, TOP, BOTTOM, SIDE, GUTTER] = [215.9, 279.4, 22, 22, 19, 7];
+const [TRIM_W, TRIM_H, TOP, BOTTOM, SIDE, GUTTER] = [215.9, 279.4, 24, 24, 38, 7];
- layout: { layoutType: 'double', gutterWidth: mm(GUTTER) },
+ layout: { layoutType: 'single' },#Code listings in boxes
For prompts with code in them, the listing boxes of Code listings and keycaps show syntax colouring with bold and italic runs.
Pitfalls
Pitfall
A nested box ignores span, placement and snapToGrid
A callout nested in another ignores its span, placement, snapToGrid and floatBarrier: it always flows inside its parent, at the parent's inner width. Nested boxes →
- A page-wide box in a two-column layout goes on the current page only when it fits with at least a couple of body lines below it. Figure 1 sits on page 1 because the abstract is set in 9.3 pt with tight padding; a longer abstract sends the figure to page 2.
Credits
- Recipe
- Ignacio Ferro
- Text
- “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models”, NeurIPS 2022, arXiv:2201.11903v6, abridged (sections 3.3, 3.4, 7 and the appendices cut, pointers to omitted figures and appendices removed); Figures 2 and 4 redrawn in code from the paper’s numbers, Figure 1 set as boxes, Table 2 reset · Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, Denny Zhou · CC BY 4.0
- Fonts
- Newsreader (SIL OFL 1.1) · IBM Plex Sans (SIL OFL 1.1) · IBM Plex Mono (SIL OFL 1.1)
Edit this write-up ↗ (opens in a new tab)Recipe folder on GitHub ↗ (opens in a new tab)


