Prompting for formats you can print

The hidden language in every answer

Chatbot answers look formatted: headings, bold, bullet points, tables. Underneath, the model is writing markdown: a plain-text convention where ## starts a heading, **word** makes bold, - starts a bullet. The chat window renders those symbols into formatting. Word, Google Docs, WhatsApp and a printed page do not. Paste the raw text and the symbols come with it.

Every teacher meets this in the first week and assumes something is broken. Nothing is. You need to tell the model where the text is going.

Say the destination

• ”Plain text, no markdown, for pasting into Word.” No symbols. Headings as capitalised lines, bold dropped, bullets as dashes or numbers. This is the safe default for anything you will paste into a document.
• ”For WhatsApp.” WhatsApp has its own conventions (a single asterisk for bold, an underscore for italic) and the model knows them if you name the app. Also implies short, and no headings.
• ”As a numbered list, one item per line, nothing else.” For anything you will cut up: questions, comments, vocabulary. The “nothing else” removes the introduction and the closing remark that every answer otherwise carries.
• ”As CSV with these columns: question, answer, marks.” For anything going into a spreadsheet. Paste into a text file, save as .csv, open in Google Sheets, Excel or LibreOffice Calc. A question bank built this way sorts and filters.
• ”As a table.” Fine if it stays in the chat or goes into a markdown-aware tool. Poor for Word. If you need a table in a document, ask for CSV and let the spreadsheet make it.

Where markdown is welcome

Some destinations do render it: Google Docs has a markdown import option, Notion and Obsidian render it natively, and many school learning platforms accept it. If you write your own materials in one of those, keep the markdown and skip the conversion. The choice is about where the text lands, not about the model.

Getting a printable page

A worksheet needs more than clean text: space to write, a name line, consistent numbering, a page that does not break mid-question. The model cannot see the page, so ask for the elements and lay them out yourself:

Give me the worksheet as plain text. Number every question. After each question write [3 lines] where students should write. Put “Name: ________ Class: ______” at the top. No introduction, no closing remark.

Paste into a document, replace [3 lines] with real space, and print. This takes under two minutes and produces something that looks like it was made rather than pasted.

For anything more designed (a poster, a card sort, a game board) the model produces the content and a free layout tool produces the page: Canva’s free tier, Google Slides, or LibreOffice Draw. Do not ask the model to design the page. It cannot see it.

Stripping symbols you already have

If you have pasted markdown into a document and it is full of symbols, do not retype. Paste it first into a free online markdown-to-text converter, or into a plain text editor and use find-and-replace: replace ** with nothing, replace ## with nothing. Thirty seconds.

Two format instructions that save a lot of editing

”Do not restate the question in the answer.” Models pad by repeating the prompt. For an answer key that habit doubles the length.

”No preamble, no summary.” Removes the Certainly! Here is your worksheet on fractions at the top and the I hope this helps your students! at the bottom, both of which you would otherwise delete forty times a term.

The honest limit

Format instructions are followed well for the first output and drift on the tenth, for the same reason register drifts: the constraint is at the top and the recent messages are closer. If a long session starts returning markdown again, restate the format in the request. And there is one format the model cannot reliably produce: precise character-level layout: aligned columns in plain text, boxes drawn from dashes and pipes. It approximates, and the approximation breaks when the font changes. For that, use the spreadsheet or the document tool. They were built for it; the model was not.