Claude Code, Codex CLI & ChatGPT: Real vs Fake Slash Commands
Most "secret slash command" lists on social media are invented. A few real ones exist, but only inside specific developer tools, and knowing which is which stops you wasting a prompt on a command that was never real.
Quick answer
Most "secret slash command" lists circulating on social media are just prompts with a slash typed in front, not real system features. Genuine native slash commands do exist, but only inside two specific developer tools: Claude Code (Anthropic's coding agent) and Codex CLI (OpenAI's separate coding agent), both of which ship real, documented commands like /init, /model and /plan. The ChatGPT app most people actually use documents almost none of its own as individual slash commands, and the "ChatGPT commands" lists that circulate online can be describing Codex CLI by mistake. You can build your own working shortcut for something like /humanize, but that's a custom instruction doing the work, not a hidden mode switching on.
Why this confusion is everywhere
An infographic claims typing /viral or /godmode into ChatGPT unlocks a hidden, developer-grade mode. It doesn't, at least not in the base chat interface. A slash isn't special to a language model by default. Unless a real system is watching for it and intercepting it before the model ever sees it, the model just reads "/godmode, write me a viral hook" as plain text and does its best guess at satisfying that instruction, the same as it would without the slash at all. That guess can be a genuinely good response, which is exactly why the myth persists: the output looks like something happened, when what actually happened is ordinary prompt-following.
Real native slash commands do exist. They're just not where most of these lists point. They live inside two specific developer tools built by Anthropic and OpenAI, each with its own command set, its own documentation, and its own rules for what triggers what.
How to check whether a command is real in your own tool
Three checks settle it for any tool, whatever its current command list looks like:
- Look for it in the tool's own documentation. A real command is something the tool's makers had to build, so it appears in their official command reference. A list that never links to that reference is a guess.
- Send the same request with and without the slash. If both answers are the same kind of answer, the slash did nothing: the model read the word after it as an ordinary instruction.
- Check where you're typing it. A command documented for a coding agent isn't necessarily available in the chat app from the same company. They're often separate products, which is where most of the mix-ups below come from.
The tables below are a sample checked in September 2026. Commands get added and renamed, so the tool's own reference is always the one to trust.
Claude Code's real slash commands
Claude Code is Anthropic's coding agent, the one that runs in a terminal, as a desktop app, on the web at claude.ai/code, as an IDE extension, and in a few other surfaces. It's a different product from the claude.ai chat interface most people open in a browser. Claude Code's full command list is documented publicly at code.claude.com/docs/en/commands and runs to well over 100 entries. A handful of the ones people actually ask about:
| Command | What it actually does |
|---|---|
/init | Explores the codebase and generates a CLAUDE.md project guide |
/rewind | Rolls back code, conversation, or both to an earlier checkpoint in the session |
/btw | Asks a quick side question without adding it to the main conversation |
/compact | Summarises the active conversation to free up context space |
/model | Shows or switches which Claude model the session is running on |
/usage | Shows token usage and plan limits for the current account (alias: /cost) |
/desktop / /mobile | Continues the current session in the desktop app, or shows a QR code for the mobile app |
/remote-control | Makes the current session controllable from the mobile app or cloud UI |
/plan | Enters plan mode for a large task before any code changes |
/doctor | Runs a setup checkup that diagnoses and fixes installation issues |
Codex CLI's real slash commands, and why they get mislabelled as "ChatGPT commands"
Codex CLI is OpenAI's own coding agent, built for the terminal and IDE, the direct counterpart to Claude Code rather than to ChatGPT. It has a genuine, separate command set, documented at developers.openai.com/
| Command | What it actually does |
|---|---|
/goal | Sets a persistent objective Codex keeps working toward across a session |
/fork | Copies the current chat into a new local session |
/memories | Chooses whether Codex uses, generates, or ignores saved memories |
/ide | Pulls your open IDE files and current selection into the next prompt |
/review | Starts a dedicated code-review pass on the current diff |
/resume | Resumes a previous Codex session |
/approve | Approves one retry after an automatic review denial |
/plan | Turns on multi-step planning before Codex starts making changes |
What's actually real inside ChatGPT itself
Strip away Codex CLI and the picture inside ChatGPT, the app people open on their phone or in a browser, gets much thinner. OpenAI's own Help Center documents typing a forward slash and "canvas" as a trigger to open the Canvas writing and coding surface, and a small handful of macOS desktop app shortcuts (quick search, image generation). It does not document a standing list of individual slash commands the way Claude Code and Codex CLI both do. Writing and code blocks in an ordinary chat now largely happen automatically as part of the response, rather than needing a separate command to trigger them. If a list claims ChatGPT itself has ten or twenty native slash commands beyond that, it's very likely describing Codex CLI without saying so.
The commands that aren't real, anywhere
Search "/godmode", "/ghost", or a string of random letters claimed as a "leaked internal command" and you won't find it in Claude Code's public documentation or Codex CLI's public documentation, the two places a real one would have to live. These circulate because a plausible-looking command name paired with an impressive result is convincing on its own, even though the result came from the model interpreting a well-written instruction, not from a hidden mode. The same applies to the most common viral list: /humanize, /tldr, /eli5, /pitfalls, /angles, /checklist and similar. None of them appear in either platform's command reference. Every one of them can still be genuinely useful, just not as a system command. They're prompt shorthand you have to set up yourself.
How to make one of these actually work
Turning /tldr or /humanize into something that reliably works takes one setup step, done once.
- Find the setting where your AI tool takes standing instructions for your chats (every chat, or every chat in one project). It's often called custom instructions, preferences or project instructions. Search the tool's own help pages for it, since where it sits changes as the tool updates.
- Write the mapping in plain language, one line per shortcut, for example: "/tldr means summarise the text above in 3 bullet points, no fluff."
- Save it, then test with a real message that starts with exactly the shortcut you defined.
- Know what you actually built: it's your own instruction firing every time that text appears, not a new capability unlocked in the model. It won't follow you to a different account, a fresh browser profile, or a chat where those instructions aren't loaded.
Related situations
- Hit an actual error message rather than a command question: check the Errors & Troubleshooting hub.
- Want the concept behind why these tools respond to structured instructions at all: What Is an AI Agent.
- Setting up something that persists across a whole project rather than a one-off shortcut: Giving Your AI Builder Project Context.
- A long chat that has started forgetting what you told it, whatever commands you use: Why Your AI Forgets or Stops.
Not affiliated with the platforms we write about
AI Build South Africa is an independent resource. We're not endorsed by, or an official partner of, Anthropic or OpenAI. We do participate in affiliate and referral programmes with some platforms named on this Site, though not these two. Command lists, pricing and product names change over time, so confirm current details against each platform's own documentation. See our Disclaimer for the full disclosure and our Terms of Service for the wider terms.
Frequently asked questions
What are Claude Code's secret commands?
There's no hidden tier to unlock. It only feels secret because almost nobody actually types a bare slash to see what's there, the way a menu only reveals itself once you open it. The complete, official reference sits at code.claude.com/docs/en/commands, covered in the table above, not in a separate leaked document.
Is /humanize a real ChatGPT or Claude command?
No. /humanize doesn't appear in Claude Code's official command list, Codex CLI's official command list, or OpenAI's documented ChatGPT features. Typing it does nothing special: the model reads the word after the slash as plain text and makes its best guess at 'humanize this'. You can turn it into a genuine shortcut yourself through custom instructions, covered below.
What's the difference between Codex CLI and ChatGPT?
Codex CLI is OpenAI's separate terminal-based coding agent, built for developers, distinct from the ChatGPT consumer app most people use in a browser or on their phone. Codex CLI has its own slash commands, including /goal, /fork and /memories; the regular ChatGPT app doesn't. Most 'ChatGPT secret commands' lists online are actually describing Codex CLI by mistake.
Are Claude Code and Codex CLI's command sets the same?
Not quite, but they overlap more than you might expect. Both document /model, /plan, /compact, /goal, /fork and /resume, and each has its own memory command (/memory in Claude Code, /memories in Codex CLI). The differences are at the edges: Claude Code's /rewind isn't in Codex CLI's documented list, and Codex CLI's /approve isn't in Claude Code's. Treat them as two separate products with their own documentation, and check the right one before assuming a command carries over.
Can I make a working /tldr or /eli5 shortcut myself?
Yes, but check it's actually working before you trust it. Type your shortcut in a fresh chat, in the same account and project where you saved it. If the AI follows your defined behaviour exactly, it fired. If it just guesses at the words instead, the instruction didn't save, or you're in a chat that isn't loading it.
Why do so many 'ChatGPT secret commands' lists get this wrong?
Because checking a primary source takes longer than copying a list that already ranks. Once one article lists Codex CLI's /goal or /fork as a 'ChatGPT command', the next ten writers copy the claim rather than opening OpenAI's own documentation, and the error compounds with each copy rather than getting caught.
Related reading
- Writing Prompts AI Coding Agents Understand: the three-part structure that actually works, so you don't need an invented command to get a better result.
- Describing a Bug So Your AI Agent Can Actually Fix It: what gets a real fix, where an invented command gets nothing.
- When to Start a New Chat Instead of Continuing: a real Claude Code habit worth having, unlike a fake slash command.