Ninety-Nine Invented Slash Commands. ChatGPT Already Has A Real List.

A ten-slide carousel, eleven entries a slide, numbered one to ninety-nine. /human: for natural writing. /eli5: to explain it simply. /premortem… no, not that one; this is a different ninety-nine from the ninety-nine the same page posted last month, with almost no overlap. Same number, different vocabulary.

And in the smallest grey type on every slide, the sentence that makes the card worth taking seriously: these are prompt prefixes, not official ChatGPT commands. Results vary and AI can be wrong, so check important answers.

That line is correct and most cards of this shape do not carry it. It is also incomplete in a way that costs you something, because ChatGPT does have official slash commands, there is a published reference for them, and one of the entries on that reference is the feature this entire carousel is hand-simulating.

What is actually on the slides

Facebook shows five of the ten images to a logged-out reader, so this is what can be verified without an account.

Type a slash in ChatGPT and something happens

The carousel’s instruction is “type one before your request.” Try that literally in the ChatGPT desktop app and the first character you press does something the card does not mention.

OpenAI publishes a slash-command reference for the composer. Its first instruction is to type / in the chat composer, then select a command from the list or keep typing to filter. A menu opens. It is populated with real commands, and — this is the part that matters later — with your enabled skills and your saved custom prompts.

So /human: is not read as a shortcut and it is not read as nothing. It opens a picker, finds no match, and you end up sending the literal text, which is exactly what the card’s grey line says will happen. The card is right. It just leaves out that you were standing in a real command system while you did it.

The two lists, side by side

Now put the card’s vocabulary next to the documented one and the relationship becomes clear.

Read down the right-hand column and the real list turns out to be about plumbing: which model is answering, how hard it is thinking, which project the chat belongs to, how much context is left, whether the chat runs locally or in the cloud, what you are being billed. There is nothing on it resembling /viral or /copywriter, and there was never going to be. These are not competing lists. They are different layers that happen to share a character.

Which is why the comparison is useful rather than a gotcha. If you learn the card’s ninety-nine and none of the real thirty, you have got very good at steering the answer and you still cannot change the model, lower the reasoning effort on a cheap question, or see how much context you have burned.

The line that turns a prefix into a command

Buried in the same reference is one sentence that makes the whole carousel redundant for anything you use twice: enabled skills also appear in the slash command list, and custom prompts appear as /prompts:<name> commands.

That is the real version of what this card is miming. A saved custom prompt is registered with the app. It shows up when you type a slash. It filters as you keep typing. It does not have to be remembered, because the menu remembers it, and it does not have to be retyped, because it is stored. If /human is genuinely something you want before every drafting request, the useful move is not to memorise a prefix; it is to save the full instruction once as a custom prompt and call it from the menu.

The difference compounds. A prefix is nine characters you type a thousand times and never improve. A saved prompt is a paragraph you can edit the moment it gets something wrong, and the edit applies to every future call. The save-once-call-in-one-word pattern is the same idea from the other direction, and it is the one thing on this topic worth actually setting up.

There is also $, which the same page gives as a way to invoke a skill explicitly rather than hoping the model picks it. Two characters, both real, neither on the card.

Slide four has three prefixes for a job the app does properly

Entries 24, 25 and 26 are /planner: build an actionable plan, /roadmap: create a roadmap to a goal, and /action: turn ideas into action steps. Three prefixes, one job, and all three produce the same kind of thing: a plan as text, in a reply, which stops existing the moment you scroll past it.

The documented list has /plan and /goal, and they are not that. /plan toggles a planning mode for multi-step work. /goal then sets what the reference calls a persistent objective that ChatGPT works toward until it finishes the task, pauses, or needs more input. While a goal is active the app shows its progress above the composer, with controls to pause it, resume it, edit the text or clear it, and you can keep steering with ordinary messages while it runs.

So the card’s three planning prefixes ask for a document about a plan. The two real commands run one. That is not a small gap in wording; it is the difference between a reply and a feature with its own progress bar.

What to use instead, prefix by prefix

Most of the ninety-nine are fine as typing habits. A few are worth promoting into something the app knows about.

The bottom two rows are the ones to notice. Two of the most useful real commands — seeing your context usage, and filing a chat into a project — have no counterpart anywhere in a ninety-nine-item list of productivity shortcuts, because a prompt prefix cannot reach them. No amount of vocabulary gets you there.

What the card gets right

Three things, and the first one is rare enough to lead with.

The grey line is honest and it is on every single slide, not just the cover: these are prompt prefixes, not official ChatGPT commands, results vary, check important answers. A card whose whole visual language is monospace slash commands on a dark terminal background had every incentive to let you believe they were real, and it tells you they are not, nine times.

The descriptions are also admirably short and concrete. “Shortest possible answer.” “Explain like I’m 5.” “Show advantages and disadvantages.” You can tell from the description alone whether the model did what was asked, which is more than can be said for a lot of prompt packs.

And the underlying instinct is sound. Having a small, fixed vocabulary you reach for without thinking genuinely does change how much you get out of a chat assistant — the quarrel here is with where that vocabulary should live, not whether you should have one. Defining your own is the version that sticks.

Four honest caveats

1. I read five of the ten slides. Facebook shows five carousel images to a logged-out reader and keeps the rest behind a login, so entries 45 to 99 are unread here. Nothing above is a claim about the full ninety-nine — the argument is about the mechanism, not the inventory, and it would hold for a list of nine or nine hundred. If any of the unseen entries collides with a real command name, that is worth knowing and I cannot tell you.
2. The slash-command reference is for the desktop app. OpenAI’s page says availability varies by environment and access, and the web and mobile composers do not necessarily offer the same set. Before building a habit on /goal or /compact, type a slash in the client you actually use and read your own menu. That check takes five seconds and beats any article.
3. “Thirty” is a count of one published table. The reference lists thirty commands at the time of writing, and also says enabled skills and saved prompts join the same menu, so your personal list is longer than thirty by design. Treat the number as the documented core, not a ceiling.
4. I have not measured whether the prefixes work. Whether /eli5: reliably produces a simpler answer than writing “explain this simply” is an empirical question about a particular model on a particular day. My argument assumes they work roughly as advertised and asks where they should live instead.

The one-line version

The card is honest that its ninety-nine are not real commands — so take the five or six you would genuinely type every week, save each one as a custom prompt so it appears in the menu as /prompts:<name>, and learn the handful of documented commands that no prefix can reach: /model, /reasoning, /status, /compact, /plan, /goal.