The LLM prompt page

A language model knows Markdown but not the studio’s dialect: it does not know that ## کتاب | ketâb | … is a vocabulary entry, that a sentence of the target language goes in […]{tl}, or that a gloss’s translation is written in italics so that the glossary can find where it ends. The LLM prompt page, /studio/prompt — the library’s LLM prompt button — holds the text that teaches it all that, ready to copy in front of your question. The model then answers with a document the studio reads as it is, typeset and verified.

That text is long, and some chatbots take less at once than the whole of it, so you choose what the model is taught. A short question about one word needs no lesson in formulas, drawings or exercises. You tick the features the document may use, and the copy explains those and no others — and still tells the model, once and in short, which marks the studio reads, so that it does not write one by accident.

1. The workflow#

The page says it in one line under the prompt:

1. Tick what the document may use, or pick a preset.
2. Copy the prompt with your question.
3. Ask your LLM: paste the lot into one message.
4. Download the .md file it writes, or copy its whole reply.
5. Back in the library, Upload .md / .zip, or Paste LLM answer.
6. Read it typeset, tag it, build the verified PDF.

A model that cannot make a file answers in the chat instead: see a file, or one fenced block, and Bringing documents in and out.

2. The page#

target language

is a menu at the top, above the boxes: the language the document will be about. It starts on the language the toolbox is set to. What it changes is what is copied: a first line stating the target (target: fa — this document is about Persian: write target: fa in the front matter), the boxes the language can use, the rules that are written differently for a script of its own than for a Latin-script language, and, under the rules, that language’s own conventions — its transliteration scheme, its reading rule, what to hyphenate.

Teach the model to write:

is the set of boxes below. A line above them says that whatever you leave unticked is still named in the prompt as reserved. Each box has its name, one line on what it lets the model write, and how many characters it adds. Ticking or unticking one changes the prompt and the size at once, and a total stays in view at the foot of the window while you tick: the prompt without your question, and how much of it is there whatever you tick. See the boxes.

The presets

tick a set in one press. See presets.

The prompt box

shows what will be copied, and its size: about N characters (about M tokens). A chatbot counts about four characters to a token in Latin script and about two in Arabic script, Chinese, Japanese and Devanagari. Past about eight thousand tokens the page says that some chatbots take less at once, and that you can ask for less or send the rules and the question in two messages.

Your question

is the box under the prompt, for what you want to know: the Persian words for hot and cold: uses, registers, opposites, compounds.

For a learner at

and length are two choices under the question, each copied as one line at the end of the prompt. See level and length. A third choice stands with them where the language has one: the scheme of its transliteration, labelled with the language’s own word for it (transliteration, rōmaji, pronunciation), usual scheme or IPA. With IPA the prompt’s first line ends · IPA, and the choice is remembered on this device, one for each language. Chinese has none yet.

The ticked boxes, the level and the length are remembered on this device, one set for every language: change the target language and the same boxes stay ticked (a box the new language cannot use is hidden until you come back to a language that can).

The prompt is written in no particular prose language: it tells the model to answer in the language of your question, so a question asked in Italian gets a document written in Italian about Persian.

3. What is always in the prompt#

Whatever is ticked, the copy has the same first and last parts:

• The task, and what the file must be: the front matter (title, subtitle, note, lang, target), ## for sections and ### for subsections and nothing deeper, and no numbers typed into titles.
• How the target language is written. A script of its own is found by the studio without a mark, and a Latin-script target — Italian, French, German, Turkish, English, Spanish — has every run of it marked […]{tl}, which the copy says for those languages only.
• One paragraph, one line. The studio joins the lines of a paragraph with a space and reads certain marks at the start of any line, so a paragraph a chatbot wrapped can be cut in the middle: a wrapped line that starts with - , * , + , > , ## , ### , a number and a full stop (a year, 1921. ), :::exercise, :::math or ::::latex starts a list, a box, a section, a numbered list or a fence there, and a line that is only a picture or a video does the same. The model is told to write each paragraph as one line, however long.
• The features are tools, not tasks. A document uses the few features it needs, never uses one to show that it can, never colours a word for the sake of it, and never describes its own markup.
• Pictures, recordings, videos and links to other documents are never the model’s to write. They are yours to add in the studio, and a model that invents one points at a file that is not there.
• The reserved marks: every mark the studio reads, with what it does, and the rule write one only for a feature you were taught; otherwise write the sentence without it. An unticked feature is therefore forbidden by name and not just left unexplained: a model that was not taught the footnote does not write x[^2] for a square, and one that was not taught the bold does not write 5 * 3 and italicise everything between the two stars. The list also names what the studio does not read at all — code fences, _x_, ~~x~~, $x$, HTML — so that a model does not write them. It grows shorter as you tick boxes: a ticked box explains its marks itself. What the studio leaves out is the same list for people.
• The answer contract, last: how to send the document (see below).

4. The boxes#

Each box explains one feature, in full, in a few hundred characters. A box that a language cannot use is not shown.

Colour inside a word teaches the model to colour only a part of one word — its stem against its ending, a root’s letters — in a way the studio still reads as one word (Colours, pronunciation and glosses shows what a reader sees). The model is told to reach for it only when a lesson is about how a word is built, to keep the word whole and on one line, and not to write it in a note written in place.

LaTeX drawings teaches the model the themes this computer has — each with the packages it adds, the ones ticked in it and the ones its own lines load, and what each package is for (xskak: chess moves and boards) — and to choose the theme by the commands a drawing needs: \ce is mhchem’s, a tikzpicture is TikZ’s, so the drawing names the theme that has the package, writes no \usepackage of its own and uses no command that none of the themes has. In the prompt that adds exercises, it also says where a drawing may stand in an exercise (a mark anywhere prose goes, a block in a prompt written prompt: | or on a Jolly card) and, when a follow-up is asked for, that a drawing in one of its steps has one colon more.

Where two features say something to each other — a colour beside a transliteration, a list inside a vocabulary entry, a formula in an exercise — the copy says it only when both are ticked, so that no box points at a rule that is not there.

5. Presets#

A preset ticks a set of boxes in one press, and shows as chosen only while the ticked boxes are exactly its set.

a short answer.

Glosses, transliteration, lists, bold and italic.

a lesson.

A short answer’s, plus vocabulary entries, tables, highlight boxes, footnotes, links and wrong and right forms. It is the set ticked when the page first opens.

a vocabulary study.

Vocabulary entries, glosses, transliteration, reading marks where the language has them, tables and lists.

a lesson with exercises.

A lesson’s, plus exercises.

all.

Every box the language can use.

none.

Only what is always in the prompt.

No preset ticks colours, colour inside a word, formulas or drawings, and only the fourth ticks exercises: tick those yourself when the document is about them. The page shows what each choice costs, in characters, before you copy it.

6. Level and length#

Two choices under the question, each copied as one line at the end of the prompt, after the answer contract:

for a learner at

— not said (the default), beginner, lower-intermediate, intermediate, upper-intermediate, advanced. The line says Write for a learner at the … level.

length

— not said (the default), short, about a page, exhaustive. The line asks to keep it short, to make it about a page long, or to be exhaustive and invent nothing.

A choice left on not said adds nothing. The prompt itself never asks for a length: how long the answer is, is what you choose.

7. A file, or one fenced block#

The prompt asks the model to answer the question by creating a Markdown file, if it can make files: download it and Upload .md / .zip. Most chatbots cannot save a file, so the prompt says what to do instead: give the whole document in one fenced block, opened with four backticks and the word markdown and closed with four backticks, and nothing else inside it. Copy the whole reply — the chatter round the block too — and Paste LLM answer: the studio takes the document out of the block, whether the fence is of three backticks or of four, with words before it and after it, and even when the document itself holds a fence of three.

An answer with no fence at all is taken as it stands, chatter and all, which is why the fence is asked for: a document that starts with Sure! Here is your document has that line in it.

8. Your own prompt#

The menu in front of copy the prompt, prompt: Parseh’s ▾, chooses which prompt this page makes: Parseh’s own, or one of your own prompts, written, changed, kept under another name and deleted with new, edit and delete beside it. The badge over the prompt says which is shown (your prompt: British spellings · Persian), and the first line of the prompt says it too (· custom: British spellings).

Added to Parseh’s

goes after Parseh’s instructions and takes the boxes exactly as Parseh’s own does: tick a box, and what it says is there, before your text.

In place of Parseh’s

is your text as you wrote it. It begins as a copy of Parseh’s own words, with the boxes’ blocks in it ({{?vocab}}…{{/vocab}}): leave them and the boxes work on it. Take them out and the prompt is copied whole: the boxes are greyed, with one sentence that says why, and what you tick changes nothing in it. The level and the length are still added, under the contract, which is Parseh’s.

Either way the answer contract is added after your text, so that the model’s answer lands the same way. Your prompts are kept in config/prompts.json and not in the library, so the library’s Backup does not carry them: export them from Settings → Your prompts if you want a copy. The editor’s Exercises ▾ → Generate with LLM… has its own boxes and does not use it: see Exercises in the editor.

9. A skill, for a chatbot that keeps them#

If your chatbot can keep a skill, copy the request for the skill beside copy the prompt puts a short request on the clipboard instead of the whole prompt: one line that names the language, the boxes you ticked, the level and the length, and then your question. The skill itself is downloaded from Settings → Skills for your chatbot and installed once in the chatbot: see Skills for your chatbot.