officeTalk syntax

What it is

officeTalk is a language for talking to AI. It has eight words, one per line, in plain English. You write in it, the page reads it back to you, and you check whether what you say you want matches what you feel you want. When it does, the page turns what you wrote into the prompt you give the AI.

Most requests to an AI go wrong before they are sent. The person is not yet sure what they want, writes down something close to it, and the AI does exactly what was written. officeTalk slows that step down and asks you about it.

The page itself is not an AI and does not call one. It runs in your browser, sends nothing anywhere and keeps nothing after you close it. You copy the prompt into an AI tool yourself.

Why an office

An AI tool behaves like an office. It acts only on what is written down, and it cannot tell what you meant beyond that. It cannot read your face or your tone.

Anyone who has had a file moved in a university office knows how this works. A file moves on its noting, the paragraph written before it goes to the person who decides, and the decision mostly follows what the noting says. So whoever writes the noting shapes the decision. officeTalk makes sure that person is you, and that you write it before the office sees anything.

A position, marked as mine

I teach design at Anant National University. The idea that AI behaves like an office is mine, from my teaching in September 2026, and so is the question the whole language is built around: what you are having to write down, does it match what you feel? It rests on a principle I hold about language and judgement: an AI should write for you only where you have agreed it can, and that agreement counts only if what you say you want and what you feel you want are the same. The sentence that says what the work is for is always yours. MATCH is that test, made into a line you write. The principle is set out in Split-Domain Cognition, in brief, under Who holds the language.

The shape of a line

Every line is a word, a colon, and plain English after it:

WANT: A list of what the submission brief asks for, marked against the list of what I have already made.

One word per line. The word can be in capitals or not; the page reads want as WANT. Capitals make the words easy to find when you read your own file back.
The colon is required. Without it the page cannot tell where the word ends, so a line like WANT a list is refused as a word it does not know.
Blank lines are ignored. Use them to separate one attempt from the next.
A line starting with # is a note to yourself. The page skips it. Use it to say what the file is about.
An empty line after a word is refused. DONE: with nothing after it says nothing, and a word that says nothing would be read back as if it did.

Why so little syntax. The language is for a person working out what they want. Every extra rule is one more thing to get wrong while you do that, so there are only as many as the page needs to read you back correctly.

The eight words

Word	What you write	In the prompt?
WANT	One sentence of what you want. Always first.	The latest one
FOR	Who it is for.	Yes
NOT	Something you do not want.	Yes
DONE	How you will know it is done.	Yes
UNSURE	Something you do not know yet about what you want.	No
HOLD	Something you are keeping unwritten.	No
MATCH	After reading it back: yes, or no and what feels different.	No
SIGN	Your name and the date.	No

WANT

WANT: A clean checklist made from the submission brief.

One sentence of what you want, in your own words. It must be the first line of the file. When what you want changes, you do not edit the old WANT: you write a new one lower down, and the latest one counts.

Why first. Writing what you want before anything else stops the request from growing out of the task. Once you start listing materials and limits, the list starts to stand in for what you want.

Why one sentence. If it takes two sentences, you want two things, and they may pull against each other. The page refuses a WANT with two sentence endings (a full stop, question mark or exclamation mark followed by a space or the end of the line), and says so. An abbreviation such as e.g. counts as an ending, so write for example instead.

Why a new line rather than an edit. Every version stays in the file, so at the end you can see how what you want changed. The note at the end shows your first WANT, your latest, and how many rewrites came between. The distance between the first and the last is often the most useful thing the file holds.

FOR

FOR: me

Who the work is for: a person or a group. "Me" is a good answer. "Everyone" is not.

Why. If you don't name a reader, you can't decide what to leave out. An office cannot guess the reader from the task. If you leave FOR out, the page asks for it.

NOT

NOT: extra items that nobody asked for

Something you do not want. One is enough; you can write several, one per line.

Why. People often find out what they want by noticing what they do not. A NOT also catches the defaults an office fills in when nobody said otherwise: a length, a tone, a conclusion you never asked for. If you leave NOT out, the page asks for one.

What the page checks. If your WANT and a NOT share a word, the page asks whether the NOT narrows what you want or pulls against it. "A checklist from the brief" with "NOT: anything not in the brief" narrows it; "a good plan" with "NOT: a plan" pulls against it. Both are fine; only you know which it is. Short common words such as that, with and what are left out of this check.

DONE

DONE: every item points to a line in the brief, and nothing is marked made that is not on my list

How you will know the work is done: something another person could check.

Why. "It should be good" cannot be checked, so you cannot tell when to stop or whether what came back is right. A DONE line is the test you will hold the work to, written before the work exists, so you set the test before you see the result. If you leave DONE out, the page asks for it.

UNSURE

UNSURE: whether I want to see only what is missing, or also what is weak

Something you do not know yet about what you want. It is allowed, and it is kept.

Why it exists. Pretending to be certain makes a WANT look finished when it is not. Writing the doubt down keeps it visible, and the note at the end lists it as still unsure.

What does not belong here. A fact you can go and check, such as whether the lab is booked or when the deadline is. You settle those by asking someone; writing them down tells you nothing about what you want.

Why it is left out of the prompt. Handed to an AI tool, a question you have not settled invites the tool to settle it for you. The doubt is yours to resolve.

HOLD

HOLD: why I am behind

Something you are keeping unwritten, named in a few words or not at all.

Why. Some of what matters most cannot go into writing without being flattened: your first rough attempts, or why you care. An office acts only on what is written. HOLD is for things you want kept out of its hands: it lets you say something is there without handing it over.

What the page does with it. It counts your HOLD lines and never repeats what they say. Nothing from a HOLD line goes into the prompt or the read-back.

MATCH

MATCH: no — a list of everything is not what I need; I need to see what is still missing from my own work

After reading your file back, you answer one question: does this match what you feel? Write MATCH: yes, or MATCH: no followed by what feels different. The page's two buttons add either line to the bottom of what you wrote.

Why this is the centre of the language. What you write down and what you want are often not the same, and the gap between them is where requests go wrong. Only you can see that gap. MATCH is where you look for it.

Why a bare no is refused. "No" on its own says something is wrong without saying what. Writing what feels different is usually how you find the next WANT: in the example above, the no is what turned a checklist of everything into a list of what is still missing.

Why a change reopens it. If you change your WANT, FOR, NOT or DONE after a MATCH, the page asks you to read it back and match again. A yes given to an earlier version says nothing about the new one.

SIGN

SIGN: (your name), 29 September 2026

Your name and the date.

Why. The person who signs answers for the work, even if someone or something else drafted part of it. When you sign, you accept what you asked for.

The questions the page asks

Once your file has no mistakes, the page asks what it leaves unclear. The questions come from fixed rules, not from an AI, so the same file always gets the same questions.

Vague words. If your latest WANT or a DONE line uses one of these words, the page asks what it looks like, so that someone else could see it: good, better, nice, interesting, engaging, impactful, meaningful, effective, clean, modern, unique, creative, powerful, beautiful. Each of these words expects the reader to recognise the result on sight. An office only has what you wrote.
A word in both a WANT and a NOT. Explained under NOT above.
A missing FOR, DONE or NOT. Each has a question of its own.

You can ignore the questions and go on.

What goes into the prompt, and what does not

The prompt is built in a fixed order from your latest WANT, then one sentence for each FOR (It is for …), each NOT (I do not want …) and each DONE (It is done when …). For the finished example it reads:

A list of what the submission brief asks for, marked against the list of what I have already made. It is for me. I do not want extra items that nobody asked for. It is done when every item points to a line in the brief, and nothing is marked made that is not on my list.

Your earlier WANTs, UNSURE, HOLD, MATCH and SIGN are left out. They are the record of how you got here, and they stay with you.

The prompt appears even before you have written MATCH: yes, with a line above it saying it is not matched yet. It is a draft until you have read it back.

Some things should not go to an AI at all, such as a statement about your work or a letter to your teacher. Write those yourself.

The rules, in one place

The first line is WANT.
A WANT is one sentence.
Every line is one of the eight words, followed by a colon and something after it.
MATCH: no needs what feels different.
Changing WANT, FOR, NOT or DONE after a MATCH means reading back and matching again.
There is no note until the latest reading is matched and signed.

What it does not do

It cannot tell whether your WANT is really yours, or whether your MATCH is honest. Those are yours, as a signature is.
It does not judge whether what you want is a good idea. It only asks whether what you wrote is what you meant.
The vague-word list is short, and a vague word that is not on it will pass.
It sends nothing and keeps nothing. Copying the prompt is the only way anything leaves the page, and only you can do that.

On the command line

officetalk.py holds the same rules for anyone who prefers a terminal. It needs Python 3 and nothing else.

python3 officetalk.py check plan.txt
python3 officetalk.py read plan.txt
python3 officetalk.py note checklist.txt

check lists mistakes and asks its questions, read reads the file back, and note prints the note that goes with the work: the first WANT, the latest, what felt different each time a reading did not match, what is still unsure, and the signature.