10 "secret commands" for ChatGPT: do they really work?

In social networks, compilations of "secret commands" for ChatGPT keep appearing. You supposedly need to add one word to the request — /human, red team, meta, or socrates — and the model will immediately start answering more intelligently, honestly, and naturally.

I have tested such commands on various tasks. Some do change the response, but there’s no magic here. These are not built-in functions of ChatGPT, nor hidden settings. The model just tries to understand what the written word means.

If you write red team, it will guess that the idea needs to be criticized. If you write ELI10, it will attempt to explain the topic more simply. But a normal instruction almost always works better than one mysterious word.

1. /human — write text more naturally

The command /human is usually placed before the request when you want a text without dry formulations and typical neural network style.

Example:

/human Write an introduction for an article about local deployment of neural networks.

Sometimes the result does become a bit simpler. But the model itself decides what to consider "human" text. It might add conversational words, emotions, or simply make sentences shorter.

In practice, it’s better to write directly what you don't like:

Write in simple human language. Don’t use bureaucratic expressions, templated introductions, or overly correct formulations. Alternate short and medium sentences. The text should sound as if the author is telling everything to a familiar person.

This way, the result is noticeably more stable. Moreover, the /human command does not guarantee passing AI detectors. These detectors often make mistakes and might label even an ordinary author's text as machine-generated.

2. Red team — find the weaknesses of the idea

Red team is one of the few commands in such compilations that can actually be useful. I use this approach when I need to check an idea before spending time and money on it.

For example:

I want to launch a website that automatically collects news from different sources. Conduct a red team analysis of this idea.

ChatGPT will try to find reasons why the project might not succeed. It will pay attention to competitors, technical issues, maintenance costs, content quality, and legal risks.

But it’s better to immediately ask the model to disagree with you out of politeness:

Act as an independent critic. Find the weaknesses of my idea, possible risks, and reasons for failure. Don’t agree with me just for the sake of politeness. After the critique, suggest how to fix the identified problems.

This kind of analysis is useful before launching a website, service, new section, or purchasing expensive equipment. It won’t replace real verification but helps notice what’s easy to overlook when you already like the idea.

3. Meta — first break down the question itself

The command meta is needed not for a direct answer but for analyzing the question itself. The model first checks whether the task is correctly formulated and if it has enough information.

Example:

Meta: is it worth buying a server for deploying local LLM?

Instead of the answer being "yes" or "no," ChatGPT may start with the budget, model sizes, speed requirements, and user count.

I usually phrase this a bit clearer:

First, break down the question. Determine what data is missing and what the decision depends on. After that, give an answer.

This technique works well with questions like "which computer is better," "should I open a website," or "which model to choose." Without additional data, it’s still impossible to give a proper answer.

4. Alt 3 — get several options

Alt 3 means that instead of one answer, three alternatives should be provided.

For example:

Come up with a title for an article about Raspberry Pi. Alt 3.

The command is simple and usually understandable. But the options often turn out to be almost the same. A few words change, but the idea remains the same.

Therefore, I prefer to immediately separate the options by style:

Suggest three different titles:1. Simple and informative.2. More emotional.3. SEO title with the main search query.

This way, the difference between the options is immediately visible. You can ask for five titles, three ways to solve an issue, or several versions of a text of different lengths. The word alt is not necessary for this at all.

5. /truth — do not speculate about unknown facts

The command /truth should make ChatGPT honestly state when it doesn't know something.

Example:

/truth What new smartphone model will be released next month?

The problem is that one word does not make the model infallible. It can still mix up dates, refer to old rumors, or present assumptions too confidently.

It's better to write clearly:

Do not speculate on facts. Distinguish confirmed information from rumors and assumptions. If data is insufficient, just say so. Don’t present a forecast as an established fact.

For news, prices, laws, characteristics of new devices, and other rapidly changing topics, it’s also important to check current sources. No code word can replace proper fact-checking.

6. ELI10 — explain complex things in simple words

ELI10 roughly translates to "explain it as if I'm ten years old."

Example:

ELI10: how does end-to-end encryption work?

The model removes most of the jargon and tries to explain the topic through an understandable example. It might compare end-to-end encryption to a closed box that only the recipient can open.

This is sufficient for a first introduction to the topic. But if I need to really understand, I ask for two explanations:

First, explain the topic in simple terms without complex terminology. Then provide a regular technical explanation.

This way, a general understanding emerges first, and then you can move on to the details.

7. /predict — describe probable scenarios

The command /predict is used for forecasts.

For example:

/predict How will the market for local neural networks develop in the next three years?

ChatGPT can gather known trends and suggest what might happen next. But this is still not a true prediction. The model doesn’t know the future and can only build versions based on available information.

Therefore, I do not ask for one precise prediction. It's better to have several scenarios immediately:

Describe three possible scenarios for the situation: basic, optimistic, and pessimistic. For each, indicate the conditions under which it may occur and what signs it may be noticed by.

This kind of response is more useful. Instead of a confident statement, you get several options and understand what each one depends on.

8. OODA — break down the problem step by step

OODA is a four-step decision-making framework:

• Observe — look at what is happening;
• Orient — understand the gathered data;
• Decide — choose a solution;
• Act — take specific action.

For example:

Visits to the new website are not growing at all. Analyze the problem using the OODA model.

First, ChatGPT will suggest collecting data: check indexing, views, positions, site speed, and user behavior. Then it will help find possible causes, choose several changes, and create a testing plan.

I like this approach because it doesn’t end with the general advice of “improve the content.” A good response should include specific actions and indicators that can later be checked for results.

9. X10 think — think deeper

The phrase x10 think supposedly makes ChatGPT think ten times deeper.

Example:

How to improve the related articles block on the website? X10 think.

The number ten does not measure anything here. The model does not start doing exactly ten times more work. It just understands that it is expected to provide more than the first obvious answer.

I would replace this command with a standard instruction:

Don’t provide the first obvious answer. Examine the problem from a technical perspective, the user’s side, and the website owner’s side. Compare several solutions, find the weaknesses of each, and choose the best option.

This way, the model understands what should be considered a deep analysis. And the result usually turns out to be more useful than after using the x10 think phrase.

10. Socrates — come to a solution through questions

The command socrates asks ChatGPT to use the Socratic method. Instead of providing a ready-made solution, the model starts asking questions.

Example:

Socrates: should I launch a new internet project?

ChatGPT might ask who the project is being created for, what problem it solves, where the first users will come from, and how much time you are willing to spend on it.

To prevent the dialogue from turning into a long questionnaire, it's better to set rules right away:

Don’t give a ready answer immediately. Ask me one question at a time, consider the previous answers, and help me come to a conclusion. When there is enough information, summarize it.

This way, it's convenient to break down ideas that are currently unclear. Sometimes after a few questions, it becomes clear that the problem was formulated incorrectly in the first place.

Are there any real secret commands in ChatGPT?

The commands from such compilations are not hidden functions of ChatGPT. There’s no special mode that activates after the word /human or /truth.

The model simply reads the request and tries to understand its meaning. The slash before the word usually doesn’t change anything either.

Sometimes a short command is enough. For example, ELI10 or red team can convey the needed task quite well. But in most cases, a precise description works more reliably.

Instead of:

/human

it’s better to write that the text should be simple, lively, and free of templated phrases.

Instead of:

/truth

it’s better to ask not to speculate on facts and honestly indicate doubts.

Instead of:

x10 think

it’s better to list from which angles the problem needs to be examined.

How to combine several techniques in one request

These approaches can be used together without issue. I just wouldn't compile them into a string of incomprehensible words.

A bad option:

meta red team alt 3 truth x10 think

ChatGPT will likely understand the general meaning, but part of the requirements may be lost.

A good option would look like this:

Analyze the idea of launching a service for local neural networks.First, determine what data is missing for assessment.Then find the weaknesses of the idea and possible reasons for failure.After this, suggest three different implementation options.For each, specify the advantages, disadvantages, and risks.Do not speculate on unknown facts and separate assumptions from confirmed information.

There are no secret words here, yet it’s clear what result is needed.

What’s the bottom line?

The majority of "secret commands" are simply short titles for well-known ways to work with information. They may change the response, but do not unlock any hidden capabilities.

I sometimes use red team, ELI10, or ask for several alternatives. But for important tasks, I still outline the requirements in ordinary words.

The clearer you explain what needs to be done, the less ChatGPT has to guess. Usually, this is what affects the quality of the response the most, rather than a slash, an English word, or an allegedly secret command.