“Build Five, Not Ninety-Nine” Is Printed On This Card, In Grey

A dense, well-made sheet went round: 99 ChatGPT shortcut commands, nine categories, numbered 1 to 99 down three columns. /DECLINE, /CUTHALF, /STEELMAN, /PREMORTEM, /TLDR. Each one with a bracket showing what you feed it and a one-line description of what comes back.

First, credit where it is due: we counted, and it really is 99. One to ninety-nine, no gaps, no repeats of a number, nine blocks that add up. That is rarer than it should be — we have audited one of these before and found a “100-prompt” sheet that was not actually 100. This one is honest about its arithmetic.

It is also honest about something much more important, in a line almost nobody will read. There are two of those lines, actually, and the whole card lives between them.

The two lines that frame the other ninety-nine

Directly under the title, in small type:

“Not built in. You create each one as a saved prompt or custom instruction — then call it by name.”

And bottom right, in the smallest and palest type on the entire sheet, next to the page edge:

“Build five, not ninety-nine.”

The headline says 99. The footer says 5. They are ninety-four apart, and the footer is the one that is right. Not out of modesty — it is right for a reason the first line sets up and the card does not have room to finish.

These are not commands. They are words you train yourself to type

We went looking for how you would actually build one, and the card’s claim holds up: ChatGPT has no feature for creating your own slash commands. Typing / opens ChatGPT’s own menu, not yours. What people mean by “custom slash commands” is writing definitions into Custom Instructions, into a Project’s instructions, or into a custom GPT — and then typing a word with a slash in front of it and relying on the model to recognise what you meant.

Three consequences follow, and all three get worse as the number grows:

• No autocomplete. A real command surfaces itself when you press a key. These do not. You have to retrieve the name from memory, unprompted, at the moment you need it.
• They are soft. The model interprets the word; it does not execute it. /CUTHALF is a suggestion that happens to start with a slash.
• They are permanent context. A definition in Custom Instructions is loaded into every conversation you ever have, whether or not you use it that day.

That third one is the quiet one. A definition like “/CUTHALF [text] — cut the length roughly in half, keep the meaning and the key points” is about 80 characters. Ninety-nine of those is somewhere near 8,000 characters of instruction riding along in front of every message you send, for the rest of the account’s life, to support a handful of commands you actually use.

Published limits on the Custom Instructions fields are hard to pin down — the figures circulating are secondary and disagree, commonly 1,500 characters on free plans and 5,000 on paid ones, and we could not reach OpenAI’s own help page to confirm. But you do not need the exact ceiling to see the shape of it. Either 99 definitions do not fit, or they fit and you pay for all of them on every message. Both answers point the same way, and the way they point is printed in grey at the bottom of the card.

Diagram: the three costs, and the one that bites

Building a shortcut costs you three different things, and they do not scale the same way. This is why the answer is five rather than fifty.

That bottom-right cell is the whole argument. A shortcut you cannot remember is not worth zero. It is worth less than zero, because the sequence is: you reach the moment, you half-remember that you built something for this, you pause to dig for the name, you fail, and then you type the request out in full. You have paid the recall tax and received nothing.

Typing “cut this in half, keep the key points” takes about four seconds. For a shortcut to beat that, it has to arrive in your head instantly, every time, without being looked up. Five names can do that. Twenty cannot. Ninety-nine is not a toolkit, it is a vocabulary test you set for yourself and then have to sit, unprepared, at random intervals, forever.

Sixty-four of the ninety-nine are near-duplicates

There is a second reason five is the right number, and it is visible on the card itself if you stop reading down the columns and start reading across them. The nine categories are named after areas of your life. The commands are not. Sort them by what they actually do and they collapse.

Look at the opening-line row in particular. /OPENERS is filed under Write & Edit, /STARTERS under Brainstorm, /HOOK under Content & Social. Three commands, three columns, one job — and when you are sitting there needing a first line, you now have to remember which of your three you built and which category your brain filed it under. The category structure, which makes the card readable, is the thing that makes the commands unmemorable.

And /TONE [text · vibe] is the one that gives the game away completely. It takes the style as an argument. It already contains /WARMER, /CONFIDENT, /HUMAN, /SIMPLIFY and /ACTIVE — five separate commands that are just /TONE with the second bracket pre-filled. That is the same collapse we found in a thirty-item card that turned out to be five operations: when a list repeats, it is usually one parameter doing the repeating.

Thirty-seven need something pasted in

One more cut, because it decides where these are any use. Read the brackets: [text], [email], [transcript], [notes], [paste], [post], [list]. We count 37 of the 99 that require an artefact you must already have and paste in. The other 62 run from a few words you type — [topic], [decision], [problem], [project].

That split matters more than the nine categories do. The 62 work on a phone, in a queue, one-handed. The 37 need a desktop, a clipboard and the thing itself already in front of you. If you are picking five, pick from the side you are actually on when the need arrives. A beautifully-built /MEETINGNOTES is no use if you never have the transcript at the moment you remember it exists.

So: which five?

Not ours — yours. But the selection rule is the useful part, and it is not “pick the five best commands.” It is: frequency × annoyance, minus how easy the thing is to type out anyway. A command earns its slot only when you hit that moment often and the longhand version is tedious enough that you keep putting it off.

Two practical notes on naming, since the names are the actual product here. Name the output, not the operation. /HOOK survives in memory because the word hook is what you want; /SUM3 does not, because at the moment of need you are thinking “shorter”, not “three”. And do not bake a number into the name. The card has six of those — /SUM3, /WINS5, /REWRITE5, /ELI10, /JOURNAL10, /IDEAS20 — and a number in the name is a number you now have to remember as well as the word. Put the count in the definition, where it belongs. The discipline of naming a size inside the prompt is right; naming it in the shortcut is not.

Four honest caveats

1. The family groupings are our reading. Somebody who writes for a living would reasonably argue that /CAPTION, /THREAD and /SCRIPT are three genuinely different jobs rather than one — we say as much in the table. The count of sixty-four moves if you draw the lines differently. The underlying point does not: this card repeats itself far more than nine tidy columns suggest.
2. The 37 / 62 split involved judgement calls. We counted [meeting] and [priorities] as things you paste; you could argue they are things you type. Recount it yourself from the card if the exact number matters to you.
3. We could not confirm the Custom Instructions character limits from OpenAI directly. Their help page returned an error for us. The figures in circulation — around 1,500 characters on free plans, 5,000 on paid — are from third-party guides that disagree on details. The character arithmetic in this piece is ours and stands on its own; the published ceiling is the part to verify in your own settings screen before you plan around it.
4. The card is right about the mechanism. We went looking for a native way to create custom slash commands in ChatGPT and did not find one, which is precisely what line two of the card says. That line is the most useful sentence on the sheet and it is set in the smallest type above the fold.

The one-line version

Ninety-nine shortcuts that you have to build, that have no autocomplete, that ride along in every conversation you ever have, and that collapse into thirteen families the moment you sort them by what they do rather than by which part of your life they belong to. The value of a shortcut is entirely in whether the name arrives in your head unprompted — and no one carries ninety-nine arbitrary labels.

The author knew. It is written on the card. Build five, not ninety-nine.