There are no secret ChatGPT codes. Here's what actually works.
You have seen the reels: type /human for human-sounding answers, EL10 to get things explained simply, mystery strings that supposedly unlock a hidden mode. Millions of views, hundreds of thousands of likes. There is just one problem: secret codes are not a thing.
AI models do not have cheat codes. They have instructions. When a made-up code sort of works, it is because the model read it like any other text and guessed what you probably meant. The guess is the weak link, and there is a better way: say the thing the code was gesturing at, in plain English.
The lines that actually work
These are instructions, not incantations, and each one exists because it fixes a specific, documented failure mode. The full list in this memo has twelve; here are the five that matter most.
Kill the guessing
Before answering, ask me up to three questions that would improve your answer.
Bad answers usually start with the AI guessing your context. Anthropic's own prompt guidance keeps coming back to the same root cause: missing specifics. This line makes the AI collect them first.
Kill the yes-man
Before you answer, argue against your own conclusion once. If the counter-argument wins, change your answer.
Models tend to agree with you and with themselves. A forced self-critique pass is the plain-English version of what the viral "killcritic" code wishes it did.
Kill the vagueness
Answer for this exact situation: [who you are, what you have, what you want]. Do not give generic advice.
One sentence of context outperforms any code ever invented. Specificity is the most boring and most powerful prompt technique on record.
Get real options, not one average
Give me three genuinely different approaches, then tell me which one you would pick and why.
The first answer is the most statistically average one. Asking for divergent options plus a recommendation gets you the edges, where the good ideas live.
Make uncertainty visible
For any claim you are not sure about, say so and tell me how to verify it.
This does not make the AI infallible. It makes the failure visible, which is what you actually need to trust the rest.
The other seven
The five above fix how the AI thinks. These seven fix how it writes and behaves, which is where most of the daily annoyance lives. Same rule as before: paste the line, get the behaviour.
Explain this to someone who has never done it before. If you must use a technical term, define it in the same sentence.
This is what the viral EL10 code was reaching for. Stating the audience beats naming a reading level, because the AI adjusts vocabulary, examples, and pace together instead of just shortening words.
Give me the answer as a checklist I can act on today. Nothing before it, nothing after it.
Left alone, models default to essays. Naming the exact shape you want is the difference between something you read and something you use.
No preamble, no summary at the end, and do not offer to help further. Just the answer.
Kills the three padding habits that make AI writing instantly recognisable. It also makes the useful part appear first instead of on line nine.
Work through this step by step first, then give your final answer separately at the end.
Reasoning before answering measurably improves accuracy on anything with multiple steps, and it lets you see where the logic went wrong instead of only that it did.
Here is an example of what good looks like: [paste one]. Match its structure and tone.
One real example beats a paragraph of adjectives. Showing beats describing, for AI exactly as it does for a new hire.
Score your own answer out of 10 against [your criteria], name the weakest part, then rewrite just that part.
A second pass on the worst section is usually a bigger gain than regenerating everything, and it costs you one line instead of another full prompt.
Before you do anything that changes a file, sends a message, or spends money, stop and show me the plan first.
The one line that matters most once your AI can act rather than just answer. It turns an agent from something you hope behaves into something you approve.
Every "secret code" that half-works is a compressed instruction. Decompress it, say it plainly, and it works fully, on every AI, forever. Magic words expire; clear instructions don't.
The same pattern runs through the other trick doing the rounds: asking ChatGPT what it "secretly" knows about you. It works, and it is not a secret either. What ChatGPT actually knows about you shows you the page it reads and why deleting a memory does not delete what it knows.
The newest version of the same folklore is the wave of free "watermark removers" that appeared after Claude started marking its text. They delete invisible characters, and the watermark is not a character. What the Claude watermark remover actually strips has the mechanism, the honest list of what genuinely comes off, and the one tool worth installing.
And in August 2026 the argument on this page got an unlikely ally. Anthropic, announcing their free training site, wrote that specific prompt behaviours are becoming less necessary as models improve, and said their education team has moved away from teaching them altogether. Claude Academy and the 4D AI fluency framework covers what they teach instead: four habits, in plain English, from a course you can take for free without an account.
Rather than copy a list, pick the behaviours that actually annoy you. Our free prompt prefix builder turns a few tick boxes into one block you paste above any prompt, in ChatGPT or Claude or anything else. Every line is plain English, for exactly the reason this memo explains.
Questions people actually ask
So the viral codes do nothing at all?
Not nothing. The AI reads them like any other text and guesses your intent, so a made-up code sometimes half-works. A plain-English instruction saying exactly what you want works every time, on every AI, with no memorizing.
Do these lines work on Claude, Gemini, and the others?
Yes. That is the point: they are instructions, not magic tokens tied to one product. Anything that can follow instructions can follow these.
Why does 'ask me questions first' help so much?
Because the biggest source of bad answers is the AI guessing what you meant. Inviting questions replaces its guess with your actual answer, before it writes a word.
What's the one line to start with?
Number one: 'Before answering, ask me up to three questions that would improve your answer.' It upgrades almost every serious request you will ever make.
Want this built for you?
We write these memos because we build this stuff every day. If you want it working in your business instead of sitting on your reading list, that is literally our job.