If you've been on TikTok or Instagram lately, you've seen them too: videos about Claude's secret codes. “Type /godmode and unlock the most powerful mode”, “put L99 at the end of the prompt and it responds like an expert”, “with Ghost you can bypass AI detectors”.
Do they really work? We tested them one by one. The short answer: no, secret codes don't exist —but behind the myth lies a useful lesson on how to get much better responses from Claude (and any AI). Let's explore it.
Index
- Where does the myth of secret codes come from?
- The 5 viral codes put to the test
- Because sometimes they seem to work
- Prompts that actually work with Claude
- Codes for Claude Code
- The real “secret code”: knowing how to use AI
- Conclusion
Where does the myth of secret codes come from?
The format is perfect for social media: "5 codes that AI doesn't want you to know," he comments to get more. It's engagement bait built on a suggestive idea: that Claude contains a hidden menu of tricks, like cheat codes in video games.
The reality is less mysterious. Claude It does not have a secret command parserThere is no locked level of ability that a magic word can unlock. Those who have systematically tested these codes have come to the same conclusion: when they "work," they work because they are compressed instructions, not secret keys.
The 5 viral codes put to the test
👉 TRY CLAUDE CODE AND COWORK FOR FREE FOR A WEEK
1. /godmode — FALSE
The promise: “unlock the most aggressive and complete response style.” The reality: there is no hidden mode. In the Claude apps, the real slash commands are the installed skills, and /godmode is not among these. At most, the model interprets the word as “be more direct and complete”—which you can explicitly ask for, with better results.
2. /L99 — FALSE (but instructive)
The promise: "Answer at the most advanced level possible." L99 isn't code: it's just text. If it sometimes produces more technical answers, it's because the model reads "level 99" and deduces that you want expert depth. Writing "Answer like a senior industry expert, assume I know the basics" does the same thing, but more reliably.
3. OODA — HALF TRUE
OODA (Observe, Orient, Decide, Act) is a real military decision-making framework, and Claude knows it. Asking him to apply it to your problem really works—but not because it's a secret code: it works because you're providing clear instructions with a structured method. This applies to OODA, SWOT, the 5 Whys, or any other well-known framework.
4. Artifacts — TRUE, BUT NOT A CODE
Artifacts are a Claude's official function: working apps, games, dashboards, and interactive documents created directly in the chat. No magic words needed: just ask "create me a dashboard for..." and Claude will build it. Passing it off as a secret code is like saying "attach files" is a hidden trick.
5. /Ghost — FALSE (and slippery)
The promise: "humanized response that outperforms AI detectors." Two problems. First: no prompts. guarantees to bypass detectors, which are notoriously unreliable in both directions. Second, presenting AI content as entirely human goes against the transparency that new European regulations require of companies. If you want a more natural tone, ask for it: "Write in a conversational tone, short sentences, no lists." This one works.
Because sometimes they seem to work
The mechanism is always the same: these “codes” are shortcuts to tell the model what you want. A language model tries to interpret whatever you write: if the context suggests “I want an expert answer,” the response adapts.
There is one documented exception, which viral videos ignore: in Claude Code (the developer tool) the word ultrathink It really does activate an extended reasoning budget. A real case—and in fact, it's no secret: it's in the official documentation.
“Secret codes” don’t unlock AI: clear instructions do.
Prompts that actually work with Claude
Here are the techniques that produce measurable improvements—no magic, just method. They're the same ones you find in Anthropic's official documentation on prompt engineering:
- Give a role and context. “"You're a marketing consultant for Italian SMEs. I need to launch a furniture e-commerce site..." — the more context you provide, the less the model has to guess.
- Show an example. Want a certain format or tone? Paste an example of ideal output and ask them to follow it. This is the technique with the best effort/result ratio.
- Define the output format. Length, structure, what to include and what not to include: “200 words max, no bullet points, close with a question.”.
- Ask for reasoning before answering. “First analyze the problem step by step, then propose the solution” — reduces errors in complex requests.
- Use explicit frameworks. OODA, SWOT, AIDA: declaring the method you want is the legitimate way to get the “code” effect.
- Iterate. The first answer is a rough draft: "more concrete," "size of the 30%," "do it from the customer's perspective." The best answers come on the second or third pass.
Codes for Claude Code: these do exist
We saw that ultrathink It's a real, documented case. And it's not the only one: Claude Code — Anthropic's terminal-based developer tool — has real commands, activated with the slash (/). But be careful: these aren't hidden tricks; they're written in the official documentation. It's the exact opposite of a "secret code." Here are the most useful ones for anyone who uses it:
/compact— summarizes the conversation to release context while retaining key decisions./clear— opens a new conversation in an empty context, preserving the memory of the project./plan—enter planning mode before a major change./rewind—returns code and conversation to a previous point (a checkpoint).
For the actual code
If you work on code, these three do the review for you:
/simplify— “clean-up only” review (reuse, simplification, efficiency) and apply fixes./code-review— check your changes for bugs and improvements; with--fixapply the corrections./security-review— analyzes changes for vulnerabilities: injections, authentication issues, data exposure.
Does your team use AI intuitively, without method?
BigFive trains people to use Claude, ChatGPT, and other AI tools with techniques that really work: practical courses built on your team's real work.
If you want to go beyond the commands and learn how to use Claude Code really well, we have collected everything in a dedicated guide: from setup to workflows, up to the keyboard shortcuts you may not know.
The real “secret code”: knowing how to use AI
The difference between those who get mediocre results and those who get excellent results from Claude, ChatGPT or Gemini is not a magic word: it is the competence of the person writing the prompt. Knowing how to provide context, examples, and evaluation criteria is a skill—and it's learned.
It's no coincidence that Europe has put it in writing: the AI Act requires companies to train those who use these tools. If you missed it, you can find it here. What does the AI Act provide for companies? —spoiler: the training requirement is already in force.
👉 TRY CLAUDE CODE AND COWORK FOR FREE FOR A WEEK
Conclusion
Claude's secret codes are a well-crafted myth: of the five viral ones, two are half-baked (OODA and Artifacts) and none are real codes. But the success of these videos speaks to something true: people feel they are not exploiting the potential of AI, and they look for shortcuts.
There's no shortcut. The good news is that the real path is short: a few prompting techniques, applied methodically, radically change the quality of what you achieve.
The perfect prompt isn't a secret: it's a clear instruction written by a knowledgeable person.
Do you want to truly harness the potential of AI in your business?
From prompt training to integrating AI into processes, BigFive supports companies and teams with customized consulting and training.
