One Of These Twenty Is A Real Keyword. It Does Not Take A Slash.

Twenty boxes, each with a monospaced name beginning with a slash: /ultrathink, /godmode, /L99, /devil, /steelman, /noyap. The headline calls them “20 secret Claude commands” and the caption says they “turn it into a completely different tool”.

Start at the bottom instead, because the card tells you what it is before anyone else can:

“These are community prompt shortcuts, not official Claude commands. Type them at the start of your message and results vary. Always check AI answers before you use them.”

That is an honest footer and it is the right one. So this piece is not going to spend time proving that nineteen of these are ordinary text — the card says so. The interesting thing is the twentieth, which is a word the software recognises, and which the documentation writes without a slash.

One of the twenty is a word the software knows

From Claude Code’s model configuration page, quoted in full:

“Include ultrathink anywhere in your prompt to request deeper reasoning on that turn without changing your session effort setting. Claude Code recognizes the keyword and adds an in-context instruction. The effort level sent to the API is unchanged. Claude Code passes other phrases such as ‘think’, ‘think hard’, and ‘think more’ through as ordinary prompt text and doesn’t recognize them as keywords.”

Three things fall out of that paragraph, and each one changes how you would use the card.

First, the slash. The documentation says include the word anywhere in your prompt. A leading slash in Claude Code means something specific: it starts a command lookup. Type a slash followed by a name that is not a command and Claude Code tells you so — it even suggests the closest match, the way claude udpate prints “Did you mean claude update?”. The one entry on this card that the software genuinely recognises is the one most likely to be broken by the formatting the card gives it.

Second, near-misses fail silently. “Think hard” is not a keyword. Nothing warns you about that; the words simply travel through as prompt text. Which is exactly what will happen to /godmode and /L99 — they are reasonable English, they will nudge the reply, and nothing anywhere will tell you they did less than you expected.

Third, even the real one is modest. It adds an instruction to the context. It does not move the effort level. The card describes /ultrathink as something that “forces Claude to reason at full depth”; the docs describe it as a request, with the API-side setting untouched.

The dial these three are reaching for already exists

Entries 1, 2 and 3 — /ultrathink, /godmode, /L99 — are three different ways of saying one thing: spend more effort on this. That thing has a control. It is called effort, it has five levels (low, medium, high, xhigh, max), and the documentation calls it “the recommended way to control thinking depth”. In Claude Code you set it with /effort, which opens a slider.

Two details make the English versions weaker than the dial. Effort “affects all tokens in the response”, thinking and tool calls included, where a sentence in your prompt affects whatever the model decides it affects. And every model carries a different default effort, so the starting point moves when you change model — we worked through that in the piece on the effort ladder.

So the practical swap is: use /effort for the session, and the bare word ultrathink inside a prompt when one particular turn deserves more. Both are documented. Neither needs a card.

Diagram: twenty names, seven jobs

Sort the twenty by what they actually ask for rather than what they are called, and the list gets much shorter — and one job gets much bigger.

Seven of twenty is worth sitting with. It is not padding — it is a signal about what people actually want from a model once they stop being impressed by it. The first thing you want is a second opinion, and the second thing you want is for that opinion to be unflattering.

But seven names with no guidance on which to pick is not help. They are genuinely different techniques, and the difference is when in the work you use them.

How to make these into actual commands

Here is the part the card leaves on the table. Everything on it can be a real slash command — you just have to write it. A custom command is a Markdown file whose body is the prompt text, dropped in .claude/commands/ for a project or ~/.claude/commands/ for yourself, and its filename becomes the command name.

So the useful move is not to memorise twenty names. It is to pick the three or four from the table above that match how you actually work, write each one as a file, and then they are commands — with autocomplete, with a slash that resolves, shared with your team if you want. We wrote about that pattern in shortcuts you define rather than features you find.

At which point the headline becomes true in a way it was not before: they are commands, and they are yours.

Four honest caveats

1. We did not test the twenty. All nineteen will produce something, because they are sensible English instructions. Whether /10x makes writing ten times sharper is not a thing anyone can check, and the card’s own “results vary” covers it.
2. The keyword finding is Claude Code specific. ultrathink is documented on Claude Code’s model-config page. In the claude.ai chat app there is no such keyword, and none of the twenty is a command there either.
3. Keywords come and go. This one is documented as of October 2026; the same page notes that other near-identical phrases are deliberately not recognised, which is the kind of detail that gets revised. Check before relying on it.
4. The card is not pretending. Its footer names these as community shortcuts and tells you to check the output. That is more disclosure than the headline needed to carry, and it is why this piece is about how to use the list rather than whether to believe it.

The one-line version

Exactly one of these twenty is a word the software recognises, the documentation writes it without the slash the card gives it, and even then it adds an instruction rather than moving the effort dial — while seven of the remaining nineteen are the same request, “tell me what is wrong”, wearing seven different names.

Use /effort for depth. Pick three of the seven critics by what they point at. Then write those three as files, and stop typing slashes at nothing.