"10 Secret Codes For ChatGPT": There Are No Codes, There Is No Secret, And Truth Mode Is A Lie

There is a fifteen-minute video doing the rounds titled '10 Secret Codes for Claude AI & ChatGPT.' The thumbnail promises secrets. The script promises 'power words' that will get you '10x better answers.' The presenter is polished, likeable, competent on camera, and repeats the phrase 'most people have no idea about these power words' early enough that you know exactly what kind of video you're in.

So let's establish the load-bearing fact before we go anywhere else, because everything downstream depends on it.

“There are no secret codes. There are no hidden modes. Nothing in that video is a feature. They are all just words, and the model is just reading them.”

The central claim is false, and it's false in a specific way that matters

The framing throughout is that these words switch the model into a different state. 'Truth mode' makes it stop lying. 'Lindy mode' makes it ignore hype. '/human' — note the slash, which is doing an enormous amount of unearned work — makes it write like a person.

None of these are parsed. There is no interpreter checking your input against a lookup table of magic strings. A large language model receives a block of text and predicts what text plausibly comes next. When you type 'truth mode,' you have not flipped a switch. You have added two tokens that statistically co-occur with a particular register of writing: blunt, contrarian, bullet-pointed, fond of the word 'brutal.' The model then produces that register. That's the whole mechanism.

That slash in '/human' is the giveaway. Slash commands are a real thing in some chat interfaces — they are implemented by the client, not the model, and they are documented. '/human' is not one of them. The slash is there because it looks like a command. It's costume jewellery.

And the 'secret' part? Search any of these ten words alongside 'ChatGPT' and you will find identical listicles going back years, on Substack, Medium, LinkedIn carousels and roughly four hundred thousand X threads. One of them, published in May 2026, is at least honest enough to open with 'these are not official hidden ChatGPT modes, they are prompt modes.' That is the accurate version of this video, and it fits in one sentence, which is a problem if your business model requires fifteen minutes.

Truth mode: the one that's genuinely harmful

This deserves its own section because it's not merely wrong, it's the kind of wrong that costs people money.

The claim is that typing 'truth mode' makes the model 'stop glazing' and 'give you the real truth, the brutally honest answer.' The demonstration is asking whether faceless short-form video is a good business for making $10,000 a month, and receiving back a sceptical answer with some arithmetic in it.

Here's what actually happened. Sycophancy in these models is not a mode. It's a training artefact — a consequence of optimising on human preference data, where humans reliably prefer being agreed with. Anthropic published research on exactly this. OpenAI shipped a GPT-4o update in April 2025 so agreeable it had to be rolled back publicly. It is baked into the weights, not sitting behind a toggle you can type.

What 'truth mode' does is change the *style* to one that performs honesty. Blunt tone. Confident verdicts. The word 'brutal' four times. You have not increased the accuracy of the output by one percentage point. You have increased its confidence, removed its hedging, and stripped out the caveats — which are, awkwardly, the only signal the model gives you that it isn't sure.

“You asked it to sound certain. It sounded certain. You mistook that for it being right.”

The failure mode is obvious once you see it. A model in 'truth mode' will hallucinate a market size, invent a conversion rate, fabricate a competitor, and deliver all three in the voice of a man who has done this for twenty years and has no time for your feelings. And because you personally typed the word 'truth,' you will believe it. You have built yourself a machine that laundering-launders confident nonsense into 'the brutal honest truth I asked for.'

In the video, the answer it produced happened to be reasonable — build an offer, don't rely on platform payouts. That is not evidence the technique works. That is evidence that 'don't expect to get rich from ad revenue on automated videos' is the median take in the training data. You'd have got it from 'is this realistic?' You'd have got it from nothing at all.

The horoscope tier

Three of the ten are cold reading with a subscription fee. They share a structure: the model is fed your own previously stated beliefs via chat memory, reflects them back with a mystical framing, and you experience the reflection as insight.

'Unlearn' — the Barnum statement generator

Pitched as a periodic exercise that surfaces the limiting beliefs holding you back. The video's own examples: unlearn that freedom comes from one dramatic leap. Unlearn that you need the perfect idea. Unlearn that you need to feel confident first. Unlearn that you need a big audience.

Read those again. Those aren't insights about you. Those are the four most common lines in the entire self-improvement corpus, and they'd apply to literally anyone who has ever typed a goal into a text box. The tell is in the video itself: 'many people who've tried it said they're always surprised by the things ChatGPT picked up.' Everyone is surprised. That's how cold reading works. Astrologers have been running this exact play since Babylon, and they didn't need a data centre.

'Future you' — advice from a character you invented

You ask the model to reply as your successful future self. It says your job isn't the enemy, your job is your investor, and you should buy your freedom in stages. Lovely. Also: a stochastic parrot wearing your own aspirations as a hand puppet, generating the median LinkedIn take with a time-travel framing bolted on for emotional effect.

As a journalling prompt, fine. Genuinely — writing to your future self is a decent reflective exercise and always has been. But it is not information. It has no access to your future. It is telling you what you already told it, in a voice engineered to make you take it seriously.

'80/20' — Pareto as a vibe

The Pareto principle is an observation about certain distributions. It is not a law of business, and the split isn't always 80/20 — Pareto was looking at Italian land ownership, not your side hustle. More importantly: the model does not have your numbers. It has no revenue data, no funnel, no churn, no cost-per-acquisition, no idea which of your activities produced anything.

So when you type '80/20' it cannot identify your highest-leverage 20%. It performs the genre. In the video, the profound output was: sell a simple AI offer to real people while your job funds the experiment. Which is fine advice. It's also the advice, unchanged, that would come back for a plumber, a poet, or a man selling artisanal gravel.

The persona trick, and the claim that isn't true

Power word four is naming a specific business influencer and having the model answer in his voice. The video says, and I'm quoting the substance closely here, that his content is all free and accessible online 'and ChatGPT can access it.'

That is not how any of this works. Absent an explicit browsing or retrieval step, the model has no access to anyone's corpus. What it has is a statistical impression of a public style — the cadence, the vocabulary, the fondness for a certain kind of formula — absorbed from second-hand summaries, transcripts, and the ten thousand people online who imitate that style badly. You are not getting his advice. You are getting an impersonation trained partly on his imitators.

And look at the formula the impersonation produced: niche × pain × offer × volume × conversion. Multiply those. Go on. What unit is the answer in? What happens when 'pain' is zero? It's dimensionally meaningless — it's a mnemonic dressed as maths, and it survived into the video precisely because it sounds rigorous while committing to nothing.

The persona technique itself is real and useful. 'Critique this as a sceptical CFO would' genuinely produces different, better-targeted output than 'critique this.' But it works because you've specified a viewpoint and a set of priorities, not because you've summoned anybody. Naming a celebrity is the weakest possible version of the technique, because you get the mannerisms and none of the judgement.

'Lindy mode' — a misused concept doing PR for platitudes

The Lindy effect, from Taleb, is a claim about non-perishable things: the longer something has already survived, the longer its remaining life expectancy. It's an argument about durability under uncertainty. It is not a synonym for 'timeless wisdom.'

Used as a prompt word, it produces exactly what you'd expect: learn to sell, learn to write, learn to teach, build trust, don't chase trends. All true. All roughly four thousand years old. All completely non-actionable at 2pm on a Tuesday when you have to decide whether to spend money on ads. Borrowing a word from a serious book does not make a fortune-cookie output serious.

Right, credit where it's due: three of them actually work

This isn't a hit piece on the technique of prompting well. It's a hit piece on calling it a secret. Three of the ten are legitimately good, and it's worth being precise about *why* they work, because the reason is the opposite of magic.

'Red team' — genuinely valuable

Asking for the strongest case against your plan is one of the highest-return things you can do with an LLM, because it inverts the sycophancy problem instead of pretending to switch it off. The model has read an enormous volume of post-mortems, security write-ups and risk assessments, so 'red team this' is a strong semantic pointer at genuinely useful territory. Caveat: the risks it names will be the generic ones. It doesn't know your specific soft underbelly. Use it as a checklist prompt, not an audit.

'ELI10' — useful, and about as secret as a doorbell

'Explain like I'm ten' compresses a wall of jargon into something you can actually check for logical holes. It's excellent for exactly the reason the video gives: if you can't follow the simple version, the complicated version was hiding something. It has also been a top-tier Reddit convention since roughly 2011 and appears in every beginner prompting guide ever written.

'Socrates' — the best one in the video by a distance

Making the model interrogate you rather than answer you is the single most underused mode of working with these things. The question the video shows — what are you actually trying to escape, the job, the ceiling, the schedule, or the feeling of wasted potential — is a good question. Not because a hidden Socrates module fired, but because you asked for questions instead of answers, and questions are where you still have to do the thinking. That's the whole trick. It is also the exact opposite of what 'truth mode' sells you.

What's actually going on: shorthand, not sorcery

Here is the honest description of all ten. They are compressed instructions. They work — to the extent they work at all — because the model has read millions of documents in which those words precede a particular kind of writing. 'Red team' is a pointer at adversarial analysis. 'ELI10' is a pointer at simplification. 'Truth mode' is a pointer at the register of bluntness, which is not the same thing as truth, and that's the entire scam.

Which means the useful skill isn't collecting incantations. It's understanding what you're actually pointing at, and pointing at it deliberately.

• →Context beats keywords. 'Red team this' gets you generic risks. 'Red team this: I have £2,000, no audience, twelve hours a week, and a competitor with a free tier' gets you something you can act on. The single word did nothing the context couldn't.
• →Put your standing instructions where they persist. Custom instructions, project instructions, a saved system prompt — that's the documented, actually-supported version of 'mode.' Retyping a magic word every session is worse than configuring it once.
• →Ask for the reasoning and the sources, not the register. 'What would have to be true for this to work, and how confident are you in each assumption?' outperforms every word in that video, and it degrades honestly when the model doesn't know.
• →Treat confident output as a red flag, not a green one. The more certain it sounds, the harder you should check. 'Truth mode' inverts this instinct, which is precisely why it's the worst item on the list.
• →Verify anything with a number, a name or a date in it. Every single time. The model's willingness to state a figure has no relationship to whether the figure exists.

The bit that made me laugh

The '/human' demonstration is a LinkedIn post about the presenter, describing her as the number one AI educator for entrepreneurs, which she then holds up as proof the output is post-ready. A creator uses an AI to generate self-promotional copy, in a video about AI tricks, in order to promote the creator. It's a closed loop. There's something almost beautiful about it — a content ouroboros, feeding on itself, humming quietly, producing nothing.

And 'sounds more human' was doing so much heavy lifting that it deserved a break. The output being self-flattering is *why* it read as good. Nobody has ever looked at a paragraph praising them and thought 'this prose is a touch mechanical.'

Verdict

This is not a con. It's worse in a boring way: it's a competent presenter repackaging a years-old listicle as forbidden knowledge, and in doing so teaching thousands of beginners a mental model of how these tools work that is flatly, structurally wrong. Believing there are hidden modes is what stops you learning the actual skill, which is specification. Believing 'truth mode' exists is what gets you to bet real money on a hallucinated number.

Keep three of the ten: red team, ELI10, Socrates. Understand that they're shorthand for instructions you could write out longhand and better. Delete the rest. And the next time a thumbnail offers you secret codes, remember that the only genuinely secret thing about prompting is that it's mostly just writing clearly, which nobody can sell you in fifteen minutes.

“There is no truth mode. There is only checking.”