Prompt prefix builder
A prompt prefix is a few plain sentences you paste above your real question that tell the model how to answer. Tick the behaviours you want below, copy the block, and put it above any prompt in ChatGPT, Claude, or Gemini. Two or three lines work better than twelve.
Do not open by praising my question or agreeing with my framing. If I am wrong, say that first. Never invent a name, number, quotation, function, or web address. If you do not know it, say you do not know it. Answer immediately. No preamble, no restating my question back to me, no opening compliment.
Why plain sentences beat “secret codes”
Every few weeks a reel goes around listing magic tokens for ChatGPT. Paste /human before your prompt, the story goes, and you unlock a hidden mode. There is no hidden mode. The model reads /human as ordinary text and guesses what you meant, which is why the same code produces a different result on Tuesday than it did on Monday.
What genuinely changes an answer is naming the behaviour you want, in words, because that is the thing these models were trained to follow. It is less exciting and it works every time. We took that apart claim by claim in the memo on ChatGPT’s secret codes.
Three that earn their place
If you only ever use three, use these. Do not open by praising my question removes the flattery that models are deliberately tuned to produce, and it is the one people notice within a single reply. Never invent a name, number, quotation, function, or web address works better than a general warning about hallucination because it names the specific categories where fabrication actually happens. And cut any sentence that would not change what I do next beats asking for a word count, which just gets padded up to.
Where it stops helping
A prefix changes how an answer is delivered. It cannot give the model information it does not have. If you need it to know your codebase, your prices, or last week’s news, no instruction will conjure that, and asking for sources will produce sources that look right rather than sources that are right. That is a job for actually giving it the material, which is a different problem and a longer memo.
Questions people actually ask
Does this work in ChatGPT, Claude, and Gemini?
Yes. Every line is an ordinary English instruction rather than a special code, so it works anywhere you can type a prompt. That is also why it keeps working after a model update, which is the part the viral code lists get wrong.
Where exactly do I paste it?
Directly above your actual question, in the same message. If you want it applied to every conversation instead of one, paste it into custom instructions in ChatGPT, or into your Claude profile settings, and you never have to think about it again.
Why does picking everything make it worse?
Instructions compete. Asking for the reasoning shown and for genuine brevity in the same breath forces the model to pick one, and it usually picks the one nearest the end. Two or three lines that target what actually annoys you beat twelve.
Is anything I type sent to you?
No. The whole tool is a single page of JavaScript running in your browser. There is no account, no server call, and no analytics on what you tick. You can turn off your network connection and it still works.
Is this the same as the secret ChatGPT codes I keep seeing?
No, and that is deliberate. Those lists claim tokens like slash-human or EL10 are hidden commands. They are not: the model is just reading them as text and guessing. These lines say what they mean, which is why they behave predictably.
This tool is the short version of a memo. There are no secret ChatGPT codes explains what is actually happening and why it works, with every claim sourced.
Want this built into your business?
We make these tools because we build this stuff every day. If you want it working inside your own company instead of sitting in a browser tab, that is literally our job.