Changelog

What’s new

New features, fixes, and improvements to llm2doc — quite a few of them started as a user report.

2026.10.11 — 2026-10-19

Tables with blank lines inside cells now export whole. Some chat exports end a cell with an empty line and continue the row underneath; only the first cells used to survive, and the rest arrived as text with stray pipes. Such rows are stitched back together in the preview and in every export.
A bare formula with absolute-value bars is recognised. A line like a sample-size bound with |S|, pasted without $ delimiters, used to be mistaken for a table row and reached Word in pieces. It is now set as one display equation, its \tag{…} number included.

2026.10.10 — 2026-10-17

A quieter look for the site. Everything is now set in one typeface, Commissioner, on a cooler, plainer page: small labels read in sentence case, headings are a little heavier, and the preview sheet uses the same serif your PDF and HTML export actually use. The layout is unchanged, and so are your documents — Word, PDF and HTML output is exactly as before.
Long headings no longer shrink the PDF. A heading too long for one line used to stretch the page and scale the whole document down to fit; headings now wrap like any other text, so every page keeps its normal size.
Long links and unbroken words wrap in the PDF too, instead of widening the page. Wide tables and long formulas still shrink to fit, so nothing is cut off.

2026.10.9 — 2026-10-16

Formulas with a missing backslash before a brace convert. A formula written \left{ … \right} instead of \left\{ … \right\} used to show as an error in the preview and reach Word as text. The brace is now repaired automatically, so the formula renders and becomes a Word equation.

2026.10.8 — 2026-10-15

Persian, Arabic, Hebrew, and Urdu read right to left in Word. Paragraphs written in a right-to-left language now open in Word and EPUB aligned and punctuated the right way, while formulas inside them stay left to right. English paragraphs in the same document are untouched.
Matrices in copied formulas keep their rows. Formulas copied through a Markdown tool now keep every row of a matrix instead of squeezing it onto one line in Word.
Word export tells you when a formula didn't make it. If a formula reaches Word as plain text instead of an equation, the export now says how many — so you know before opening the file, and we see the pattern to fix it.
Formulas copied through Markdown tools convert in every form. The same fixes now cover formulas written with dollar signs (as Gemini, Claude, and DeepSeek write them), sets like \{a, b\}, and copies where every backslash was doubled — all become real equations in the preview and in Word.

2026.10.7 — 2026-10-13

Formulas copied through Markdown tools reach Word as equations. Some copy-as-Markdown tools add an extra backslash to formula markers like \[ … \], and most of those formulas reached Word as raw code. Now every one becomes a real Word equation.
LaTeX examples in code blocks stay exactly as written. If an answer shows how to type a formula inside a code block, the code is no longer turned into math.
Equations with extra line spacing keep their layout. Multi-line equations that add space between lines (like \\[2pt]) now convert correctly.

2026.10.6 — 2026-10-10

Formulas from scanned or OCR'd documents reach Word as equations. Text recognised from a scan often writes formulas with spaces inside the dollar signs, like $ x^{2} $ . The preview showed them as math, but the Word file got the raw code; now they become real Word equations too.
Pictures from scanned documents appear in Word. OCR tools wrap each figure in a bit of HTML that Word export used to drop; those figures now come through, in PowerPoint and EPUB too.
Pictures linked from the web are fetched more carefully. Word, PowerPoint, and EPUB exports embed linked pictures up to 5 MB each (8 MB per document) and show a picture's description when it can't be fetched — and documents with diagrams convert using less memory.

2026.10.5 — 2026-10-09

Oversized share links are turned away in seconds. A ChatGPT or Gemini share link to a conversation too long to import now gets its "copy the part you need" answer as soon as the page passes the limit, instead of after the whole page has downloaded — and the converter stays quick for everyone else while it happens.
Word, PowerPoint, and EPUB handle documents up to about 200 pages. A longer document now gets a clear message right away instead of a slow attempt — PDF and HTML export still work at any length, or split it into parts. Very long chat conversations import their most recent ~200 pages, so an imported conversation always converts to Word.

2026.10.4 — 2026-10-06

Formulas in PDFs print in proper math type. Equations in a PDF now use the same math fonts as the preview: ℝ, script and fraktur letters keep their shapes, and spacing around signs like "=" matches what you see on screen, with no swallowed space after a formula.
Footnote mode leaves units alone. With "Citations as footnotes" on, squared and cubed units such as m² or a power like x² stay as they are instead of turning into footnote marks, footnote numbers stay attached to their word, and formulas are never touched.

2026.10.3 — 2026-10-05

Every formula in a Gemini copy reaches Word as an equation. When a copied answer escapes its math (every backslash doubled), formulas on their own line already converted — now the ones inside sentences do too, instead of arriving as raw code or with stray underscores.
An open tab tells you when llm2doc was updated. If you exported from a tab opened before an update, the export used the older version. Now Download offers a quick reload first — your text and settings stay — so every export gets the latest fixes.
Steadier under heavy use. Importing very long chat conversations and converting to Word now take turns on our server and run within a tighter memory budget, so one big job no longer slows things down for everyone else. Share pages up to 15 MB import directly; for a longer conversation, copy the part you need and paste it.

2026.10.2 — 2026-10-04

Sharper PDFs, however long the document. Every page of a PDF is now drawn at full resolution. Longer documents used to come out at a lower resolution that looked soft when you zoomed in; now page 50 is as crisp as page 1.
Long PDFs export in Safari. Long documents now export to PDF in Safari too, and long exports are faster in every browser.
Long PDFs export more than twice as fast. Each page now takes far less work to build, so long documents finish in a fraction of the time — at the new, sharper resolution.
Prices written as "10$" stay prices. Two amounts with the dollar sign after the number could be read as a formula, squeezing the text between them into one long line — and shrinking a whole PDF to fit it. They now stay ordinary text, and the PDF keeps its normal size.
A share link in the editor always becomes the conversation. If a ChatGPT or Gemini share link lands in the editor some other way than a normal paste — from a phone keyboard's clipboard suggestion, dragged in, or typed — it now imports, and Download never turns the bare link into a one-line document.
Very long chat conversations import instead of being turned away. A ChatGPT or Gemini share link to a conversation too long to convert in one go now imports its most recent messages, with a note at the top saying how many of the total it holds.

2026.10.1 — 2026-10-03

Mermaid diagrams arrive in Word as diagrams. Flowcharts, sequence diagrams, and the rest used to reach Word, EPUB, and PowerPoint as their source code. Each one is now drawn in your browser and embedded as a picture, at the size it has on the page.
Equations typed as text in a Word file survive the trip to Markdown. When a .docx has formulas written out as $ … $, \[ … \], or $$ … $$ instead of built with Word's equation editor, uploading it now brings them back as real math — in sentences, in table cells, and on their own lines. Before, the brackets were lost and the formula came out broken.
Formulas that lost their delimiters convert inside sentences. An answer copied without its $ signs — "Margin: \frac{12}{16} \times 100 = 75% of sales" — now shows the formula as math in the preview, Word, and PDF, while the words around it stay text.
Tables pasted from Word or Google Docs arrive in Word. A table copied from Word or Google Docs now lands in the editor as a clean, editable table instead of a block of Office formatting code, and every table in your document — including ones written as HTML — comes out in Word and PowerPoint as a real table. Before, such tables could go missing from the Word file.

2026.9.8 — 2026-09-26

Long documents export to PDF with every page. A very long document — an imported conversation, a big report — could come out as a PDF of blank pages. The PDF is now drawn in pieces your browser can always handle, so every page has its content. If a PDF ever comes out empty anyway, you get a message instead of a blank file.
Feedback names the right format. If you export twice and then answer "Did it come out right?", your answer now counts for the export you just did, not the first one.

2026.9.7 — 2026-09-23

A share link pasted on any landing page imports the conversation. The paste boxes on the format pages used to hand a link to the converter as text, so exporting gave you a document containing only the link. They now import ChatGPT and Gemini share links the way the converter itself does.
A link we cannot import says so. Paste a link to another chat service or any other site, on its own, and the page tells you to open it, copy the reply, and paste the text — instead of quietly converting a URL.
Very large ChatGPT conversations import again. Some big share pages failed with an unexplained error. They now import.
Equations copied with doubled backslashes render. Some copy buttons escape the reply, so formulas arrive as \\frac with escaped brackets. They used to show as broken text in the preview and in Word; they are now proper equations.

2026.9.6 — 2026-09-21

Open in Google Docs is reliable now. On Safari, and on any browser with a long or equation-heavy document, the copy could be refused and the export failed however many times you tried. The copy is now set up the instant you click, so it goes through. Copy for Notion got the same treatment.
It works from Editor view too. Exporting to Google Docs with the preview hidden used to ask you to switch views first. The preview now opens by itself and the export carries on.
When the browser blocks the copy, a second click gets it done. Some browsers, in-app ones especially, refuse clipboard access outright. The message now says so plainly, and clicking the button again copies the document by another route that those browsers allow.

2026.9.5 — 2026-09-20

Formatting buttons and shortcuts are steadier. Bold, Italic, and the other toolbar actions now always act on the text as it is right now. Pressed the instant after an edit, a shortcut could do nothing, and a selection from an earlier action could come back a moment later. Both are fixed.
An old guide address finds its page. The previous HTML to Markdown guide link now opens the HTML to Markdown page instead of the guides list.
LaTeX pasted on its own now converts. An equation, align, gather, multline, or eqnarray block copied from a .tex file, or a line that is just a formula, used to show as raw code because it had no $ around it. It now renders as an equation in the preview and arrives in Word as editable math. Environments the preview could not draw before (multline, eqnarray, anything with a \label) render too.

2026.9.4 — 2026-09-08

Print fills the page. The Print button (and Ctrl/Cmd+P in the viewer) now prints at your document setup's paper size and margins, with body text at its true point size — no more narrow column down the middle of the page. The cover page and the contents list each take a page of their own, headings stay with the text below them, tables and code blocks avoid splitting across pages, and when page numbers are on in the setup they appear in the footer (in browsers that support page footers). The table → Excel chip no longer prints, and the PDF you save is named after your document's title rather than the site's.
Word: equations with = on its own line. Chat models often write a display formula with the = sign alone on a line. Word exports turned those into a large heading of raw LaTeX with the rest of the formula as text; they are now proper editable equations, as they already were in the preview, PDF, and HTML. EPUB and Slides get the same fix.
Recent conversions, right under the converter. The documents you exported used to hide in a collapsed list at the foot of the page. They now sit directly below the converter as small cards — the document's title, the formats it went out as, and when — one card per document however many times you exported it. Click a card to load the document back into the editor; hover and × to forget one. Still stored only in your browser.

2026.9.3 — 2026-09-07

Word and HTML to Markdown. The reverse direction: drop a .docx (or .odt, .rtf, .html) on the editor or the Upload button and it comes back as Markdown — headings, lists, pipe tables, footnotes, images inlined, and every Word equation as LaTeX — ready for a ChatGPT or Claude prompt or for another export. Two new pages, Word to Markdown and HTML to Markdown, take files and pasted HTML. And formatted text pasted into the editor — from Word, Google Docs, an email, a web page — now lands as Markdown instead of plain text.
Gemini share imports keep their sources. A conversation imported from a Gemini share link now carries each Google Search source right after the sentence it supports, so research mode can turn them into footnotes.
Bracketed formulas inside a bullet stay inline. A \[ … \] formula with text around it on the same line — "Distance covered: [ d = vt ] where…" — now renders as inline math instead of a display block that cut the bullet in two. Word files imported back to Markdown also lose the escaped brackets and three-space list markers the conversion used to leave behind.
A newer conversion engine. Word, EPUB, and slides exports and the Word import now run on Pandoc 3.11. Two things you will notice: the Word contents list gets its page numbers as soon as Word asks to update fields on opening (no more F9), and Word files imported to Markdown keep their horizontal rules.
Contents jumps land every time. Clicking a heading in the Contents menu scrolls the preview all the way to it in the newest Chrome, where the jump used to stop short.
Cover and contents behave on real documents. A document without headings no longer gets an empty contents page — and Word no longer complains about it when the file opens. The cover title falls back to the document's first line when there is no H1. In the PDF, a formula wider than the page used to push the cover and contents into the middle of a page; they keep whole pages now. The setup panel shows how many headings the contents list will use and warns when there are none.
Two more format pages. Markdown to HTML and Markdown to EPUB, and the home page's format chips link to them; the Slides chip opens the converter with slides pre-selected. EPUB books now take their title from the document's first heading instead of "Document".
A home page you can scan. Below the tool, the paragraphs are gone: an In → Out strip shows what you can bring (a copied reply, a ChatGPT or Gemini share link, a Word, HTML, or Markdown file) and what you get, six one-line features link their pages, and the answer to "how do I convert a ChatGPT reply to Word" is two sentences.
Pictures no longer flood the editor. An image that arrives as base64 — from a Word import or a pasted picture — now shows in the editor as a short image:1 reference. The preview, every export, the share link, and the Markdown download still carry the picture itself.
A footer you can scan. The one long row of links is now four columns — Markdown to, From a chat, Back to Markdown, Site — and every page, including the finance and equations pages, is linked from it.
Two more ways to a PDF. New pages for Gemini to PDF — paste a public share link for the whole chat, or one answer — and Claude to PDF, alongside the ChatGPT one.
Choose how the Word contents list gets its numbers. By default Word fills them in when the file opens and asks once; a new switch in Document setup, "On F9", keeps the list pre-filled without numbers and Word opens the file without asking.
Section titles survive the trip through Word. A Word file or a formatted paste with no heading styles — an AI answer whose titles were only bold, or numbered "1. Scope" lines — now gets those titles back as headings, so the contents list, the Contents navigator, and the cover title work on it.
Wide equations fit the page. A display equation wider than the column now shrinks to fit — in the preview and in the PDF, where one long line used to shrink every page — and scrolls only when it would have to drop below 60%.
One-line $$…$$ is display math. Written alone on its line, it now renders centered and full-size in the preview, HTML, and PDF, the way Word already showed it.
Wide equations copy whole. Copy formatted no longer clips the right edge of a display equation wider than the page.
Dollar signs inside code stay put. A $1 placeholder in a SQL or shell code block, or in inline code, no longer picks up the backslash the currency guard adds to prices in prose.
Fuller examples. The Load example menu's documents are now real ones — a 500-word article with footnotes and a task list, four tables with alignment and currency, rate limiting in three languages, a migration plan, formulas from Maxwell to matrices, a full investment appraisal — plus a new Research brief with numbered sources that shows research mode off.
Pictures show in the preview. Images embedded in the document itself — a Word import's pictures, an image pasted with formatted text — rendered as broken icons in the preview even though every export carried them. They render now.
Copy formatted round-trips its equations. The equation pictures it puts on the clipboard carry the formula as their description, and pasting them back into the converter — directly, or via Word or Google Docs — gives you the equation again instead of a picture.
The Upload button explains what it takes. It always accepted Word, OpenDocument, RTF, and HTML files alongside Markdown; hovering it now lists them, and the editor's placeholder mentions dropping them too.

2026.9.2 — 2026-09-06

Copy the document without the asterisks. A new Copy menu above the preview offers two options: Copy formatted puts the rendered document on the clipboard for Word, Outlook, Gmail, or Google Docs — headings, lists, tables, and code arrive as formatting (equations as images; use Word export for editable math). Copy plain text gives clean text with every ** and ## removed, lists and tables kept readable, and tables tab-separated so they paste into Excel as cells.
Excel export. Pick Excel (tables) in the format picker and every table in your document becomes a sheet in one .xlsx — named after the heading above it, header row in bold, and numbers arriving as numbers so formulas work on them. Built in your browser; nothing uploads. Every table in the preview also gets a small Excel chip when you hover it: download that table alone, or all of them at once. Tables written as HTML — what assistants use for merged cells and two-row headers — export too, with a caption used as the sheet name.
Research mode: citations become real footnotes. Paste a Perplexity or Deep Research answer with [1] markers and a Sources list, or a ChatGPT web-search answer with its source links, and a hint offers to turn the citations into footnotes — proper Word footnotes at the foot of the page, a footnotes section in PDF and HTML. Sources nobody cites stay in the list, your text in the editor stays exactly as pasted, and the switch lives in Document setup as "Citations as footnotes". Conversations imported from a ChatGPT share link now keep their web-search sources too.
Document setup. A new sliders button above the preview opens a panel for the things a document is usually asked to have: A4 or Letter, narrow / normal / wide margins, a sans or serif body and headings at 10–12 pt, tight / normal / loose line spacing, page numbers, a contents list built from your headings, and a cover page with title, author, and date. The preview shows it all as you change it. Word gets all of it — the contents is a real Word field that fills in page numbers when you update fields (F9) — and PDF gets the page size, margins, fonts, cover, contents, and numbers. You can also hand in your own Word template (.docx): its styles, page setup, and headers then shape every Word export, with a switch to apply the panel's page and fonts over it. Choices are remembered on your device.
Import a ChatGPT or Gemini conversation from its share link. Paste a chatgpt.com/share or gemini.google.com/share link into the editor (or use the new ChatGPT link to Word and Gemini link to Word pages) and the whole conversation loads: the title as a heading, your prompts as quoted blocks, every answer verbatim with its tables, code, and equations — ready for Word, PDF, or any other format. Gemini's short share.gemini.google links work too. Our server reads the public page once and keeps nothing. Claude links are recognized and get a clear "paste instead" message, since Claude's share pages can't be read by a server yet.
New pages: a Markdown table to Excel converter page, a guide on pasting ChatGPT into Word with formatting (no asterisks, no hashes), and FAQ answers for exporting tables to Excel and for ChatGPT's “conversation is too large to export” message.
The LinkedIn, X, and email placeholders are gone from the format picker.
Equations arrive whole in Google Docs and in Copy formatted. Tall formulas — fractions, integrals with limits, square roots — used to be cropped to a thin strip when copied; every equation image now covers the full formula, and its description is the LaTeX source.

2026.9.1 — 2026-09-02

Paste and convert on the page you land on. The Markdown-to-Word, Markdown-to-PDF, and live-preview pages, and the ChatGPT-to-Word guide, now have a paste box at the top: paste, click once, and the workspace opens with your document loaded and the format already chosen. Dropping a file still works too.
The converter pages explain what they do. Markdown-to-Word and Markdown-to-PDF gained a plain-language walkthrough of what happens to your text, a table of what each Markdown construct becomes, and fuller answers to the common questions.
Guides link to every other guide, and the equations and finance pages now appear on the guides index.
The LaTeX cheat sheet is lighter and reads cleanly for screen readers: each formula's LaTeX source is its accessible name.
The homepage says, in one paragraph, what llm2doc is and what it exports.
Three new converter pages, each with a paste box: LaTeX to Word (equations arrive as native, editable Word math), ChatGPT to PDF (a clean A4 PDF of the answer, built in your browser), and Markdown to Google Docs.

2026.8.8 — 2026-08-26

A heads-up when your AI sent code instead of a document. If an assistant answers with a script that generates a document rather than the document itself, the workspace now points that out before you export — with what to ask for instead. Dismissible, and it never appears on ordinary documents.

2026.8.7 — 2026-08-23

Sturdier conversions for very large documents. Each conversion now runs with its own memory ceiling and a 60-second time limit — a document too large to convert fails quickly with a clear message, and never affects anyone else's exports.
Oversized uploads are rejected immediately on arrival.
Export buttons respond instantly. The working indicator now appears the moment you click, before the conversion starts — noticeable on large documents, where the click itself used to stall briefly.
Terminal-style tables become real tables. Answers copied from CLI tools draw tables with ┌─┬─┐ box characters; those now convert to proper tables in the preview and every export. Flowcharts and file trees are left as-is. (From two user reports.)
Long two-column tables survive the trip to Word. Line breaks inside table cells (<br>) no longer split the row — a 200-row table now arrives in Word with all 200 rows. (From a user report.)
Cleaner PDF page breaks. A section heading no longer ends up stranded at the bottom of a page with its content on the next one, and a very tall block continues across the break instead of leaving a large gap. (From a user report.)

2026.8.6 — 2026-08-17

This page. llm2doc now keeps a public, versioned changelog — every user-facing change lands here with a date, including the user reports that drove it.
Smoother under heavy load. Simultaneous exports (a whole classroom clicking Download at once) now queue one at a time, so every conversion stays reliable during busy moments.
Documents over 2 MB of markdown are now politely declined with a clear message instead of risking a failed conversion. No real document has ever come close to the limit.
Two more paste styles render correctly (thanks to two user reports): display math with $$ on the same line as multi-line content (previously showed raw LaTeX in the preview), and Word-style pastes where inline variables arrive as bare parens — (v) = velocity and (F_g) now become real math while "(a) or (b)" stays prose.

2026.8.5 — 2026-08-13

Typing is ~10× more responsive on long documents. The preview no longer re-renders on every keystroke; on a 20 KB pasted answer, keystroke response went from ~416 ms to ~40 ms.

2026.8.4 — 2026-08-12

Word equations no longer silently degrade. If the equation service doesn't respond, the converter retries automatically; if it still fails, you get a clear warning that equations came out as plain text — instead of a broken file presented as a success. (From two user reports.)
Failed conversions now tell us why they failed, so problems get fixed from the first occurrence instead of the fifth.

2026.8.3 — 2026-08-10

PDF pages no longer slice text in half at page breaks. Page cuts are now verified against the rendered pixels, so a line of text can't end up split across two pages. (From a user report — thank you for attaching the document.)
Inline math that opens with a digit — $2 < \text{min\_sup}$ — now renders instead of being mistaken for a price.

2026.8.2 — 2026-08-04

Fixed links being destroyed when their URL slug looked like a formula (underscores read as subscripts).
Back-to-back links now get a separating space instead of running together as one word.
API conversions now go through the same markdown cleanup as the site — ChatGPT math delimiters and currency are handled identically.

2026.8.1 — 2026-08-01

Word and PowerPoint keep styled callout boxes instead of silently dropping their content. (From a user report.)
Fixed blank multi-hundred-MB PDFs some users got — PDFs are now properly compressed as well (a 9 MB page became 0.2 MB).
Word documents match the site's look: serif ink headings instead of Word's default blue.
One code-block color scheme everywhere — the preview's editorial palette now carries into Word, PDF, HTML, EPUB, and slides.
PowerPoint export reworked: widescreen 16:9 slides with full-width content (was a narrow side column), every slide gets a title, and the format picker now explains what PPTX export can and can't do.
Fixed a bug where a PowerPoint file could be rejected by PowerPoint as corrupt.

2026.7.9 — 2026-07-30

Feedback can now include a copy of your markdown (on by default, opt-out) so reported conversion problems can actually be reproduced and fixed.

2026.7.8 — 2026-07-29

Better on phones: readable 16 px text and comfortable tap targets.
Faster first paint — the demo document renders with the page instead of after it.
Negative feedback now reaches us as a GitHub issue, message included — the fastest path from "this looks wrong" to a fix.
New About and Terms pages, and a proper 404 page.
The /markdown-viewer page now embeds a working viewer — drop a .md file straight onto it.

2026.7.7 — 2026-07-19

Share, Contents, and Print collapsed into one quiet toolbar above the preview.
?to=pdf-style deep links pre-select any format, with aliases.
Fixed Google Docs export leaving the clipboard empty; fixed printing emitting a blank extra page.
New favicon: the wordmark's arrow.

2026.7.6 — 2026-07-18

Share links: the Share button packs your whole document into the link itself — nothing stored on any server; opening the link renders the document.
Formatting toolbar v2: buttons toggle, work per-line on multi-line selections, and formatting from a preview selection maps back into the editor. Added strikethrough, task lists, images, a status bar, and native undo.
New /markdown-to-pdf landing.

2026.7.5 — 2026-07-17

The workspace became a real viewer/editor: Split, Editor, and Preview layouts plus a full-window Focus mode (Esc leaves it), and Print.
Formatting toolbar with ⌘B / ⌘I / ⌘K shortcuts.
Mermaid diagrams render in the preview (and in PDF exports).
Contents dropdown: jump to any heading in a long document.
New guide: how to open a .md file.

2026.7.4 — 2026-07-07

👍/👎 feedback prompt after each export, with reason chips.

2026.7.3 — 2026-07-06

Four new export formats: open in Google Docs, EPUB, PowerPoint slides, and Copy-for-Notion.
Drag-and-drop a .md file straight onto the editor.
Currency amounts like $18K no longer render as broken red math.

2026.7.2 — 2026-07-05

Upload .md files — convert files, not just pastes.
Click any equation in the preview to copy it as Word-ready math — paste into Word and it becomes a native editable equation.
New guides for Grok and DeepSeek; new landing pages for markdown-to-Word and ChatGPT equations; finance worked example.
MCP server: AI agents can call llm2doc's conversion directly.

2026.7.1 — 2026-07-04

Paper & Ink redesign — the editorial look the whole product now wears: single accent color, Newsreader serif, calm layouts.
First visit shows a worked demo document instead of an empty box.
Rewrote the LaTeX guide as a cookbook with rendered examples; added ChatGPT export troubleshooting.