Decoded
The ChatGPT Prompt That Decodes Her Texts (and Where It Falls Short)
A real prompt you can copy and use for free, right now — and an honest account of the four things a general chatbot gets wrong when you hand it her message at 1am.
Yes, ChatGPT can do this well, and the prompt below is free. Copy it, use it tonight, and close this tab if that’s all you came for. But it has four specific weaknesses that matter when you’re anxious, and you should know them.
We make an app that does this. We’re still going to give you the free version first, because a page that hides the working answer behind a download button isn’t worth reading. If the prompt is enough for you, that’s a good outcome. Most nights, it is.
The prompt
Paste this into ChatGPT — it works just as well in Claude or Gemini — then fill in the three fields at the bottom and send. Don’t add anything else to it.
Copy from here.
You are helping me understand a text message I received. I will give you the exact message, unedited, and where I am in the relationship.
Do not flatter me and do not automatically take my side. If my framing of the situation is unfair to her, say so. If the message is ordinary and means exactly what it says, say that plainly instead of inventing a hidden meaning — most messages are not code.
Answer in exactly these five parts, with these headings:
- WHAT SHE LIKELY MEANS. The most likely plain reading. If two or three genuinely different readings are possible, list them, say which is most likely, and say what would tell them apart.
- WHAT SHE MIGHT BE FEELING. The emotion that could be driving this — or “probably nothing much” if that is the honest answer.
- HOW SERIOUS IS THIS. One word: GREEN (straightforward, stop overthinking it), YELLOW (fine, but the reply needs care), or RED (a real problem that needs a conversation, not a clever text). Then one sentence on why.
- WHAT I COULD SEND BACK. Two or three replies I could actually send. Plain and honest. No games, nothing designed to pressure her, nothing engineered to make her chase me.
- WHAT NOT TO SAY. The reply that would most likely make this worse, and why.
If the honest answer is “just ask her what she meant,” say that instead of guessing.
Rules: use only the message text and the context I give you. Do not invent history you were not told. If you are unsure, say you are unsure, and say what you would need to know.
Message (exact, unedited): “…”
Relationship stage: dating / together / married
What was happening right before: one factual sentence, or “nothing”
That’s the whole free version. It is not the prompt our app runs on — that one is ours, and it’s the only thing we actually built — but it is put together the same way, and it will get you a usable read.
How to use it
The prompt does about half the work. The other half is what you feed it, and this is where most people quietly sabotage themselves.
- Paste her exact words. Not your summary of them. Not a tidied-up version. The punctuation is data — “ok” and “ok.” and “Ok!” are three different messages, and if you retype it from memory you will smooth out the very thing you’re trying to read.
- Give it the relationship stage. “Sure” from someone you’ve been seeing for three weeks and “sure” from your wife of nine years are not the same sentence.
- Don’t lead the witness. If you open with “my girlfriend is being passive-aggressive again, look at this,” you’ve already told it the answer. It will hand your theory back to you with better vocabulary.
- One factual sentence of context. Not a case for the prosecution. “I cancelled on her yesterday” is context. “I cancelled once and she’s been impossible ever since” is an argument, and you’ll get an argument back.
- Ask once. If you keep regenerating until an answer feels good, you’re not analysing her message any more — you’re shopping for reassurance, and you will always find some.
What good output looks like
Take this one. She writes: “you don’t have to come if you don’t want to.” You’ve been together two years. You said earlier that you were tired.
A good answer does not tell you what she means. It tells you what she probably means, and what would settle it. Something like: the most likely reading is that she does want you there and is giving you an exit she hopes you won’t take — she may be feeling unimportant rather than angry. The second reading is that she means it literally and is being generous about your tiredness. The way to tell the difference is not to analyse the sentence harder. It’s to answer the actual question underneath it, which is do you want to be there.
Yellow, not red. And the reply that makes it worse is “ok, I won’t come then” — not because it’s a trap, but because it answers the words and ignores the person.
If the output reads like a horoscope, throw it away. “She may be feeling a complex mix of emotions” is not a reading, it’s a hedge. A good answer commits to a most-likely meaning, names what would disconfirm it, and is willing to say “this message is fine, go to sleep.”
Four things a general chatbot gets wrong here
1. It agrees with you
This is the big one. A general assistant is built to be agreeable, and it will follow your lead almost anywhere. Tell it she was being unreasonable and it will explain why she was unreasonable. Tell it you were the problem and it will explain, with equal confidence, why you were. Both answers will sound like analysis. Neither one is.
The instruction in the prompt above — do not automatically take my side — helps, and it is the single most important line in it. But it does not make the model neutral. If you frame her badly, you will get a reading of a woman who doesn’t exist. Be careful what you tell it before you ask.
2. Its severity scale drifts
Ask it twice and you can get two different answers about how bad this is. The same message can come back “mild irritation, nothing serious” on the first run and “significant hurt” on the second, because nothing in a chat window holds a general model to a fixed scale. That’s fine when you’re brainstorming. It’s not fine when the thing you actually need is a straight answer to “is this bad?”
3. It won’t tell you what not to say unless you make it
Ask a chatbot for advice and it will give you things to do. It almost never volunteers the reply that would blow this up — and in a text argument, the damage is usually one message long. That’s why part five is in the prompt. Take it out and you lose the most useful thing on the page.
4. It’s 200 words you have to remember at midnight
This is the unglamorous one, and it’s the reason most people stop. The prompt works. It also requires you, at 1am, with your stomach in a knot, to find it, paste it, fill in three fields correctly, and resist the urge to editorialise. That’s a lot of executive function to ask of a man who has just read “we need to talk” — and if you’re in that specific hole right now, read this instead before you paste anything anywhere.
What it can’t do — and neither can we
Here is the part we are not supposed to write.
ChatGPT does not know her. It has never met her. It doesn’t know that she goes quiet when she’s tired, or that her mother is ill, or that “k” is just how she types on the train. It is reading a sentence out of context and giving you a statistically sensible guess about what a person might mean by it.
Decoded doesn’t know her either. Same limitation, exactly. We read one message, we don’t see your thread, and we can’t see her face. We don’t publish an accuracy percentage because we don’t have one and nobody in this category honestly does — including the apps that quote you a number.
Which means the last call is always yours, and it always was. A decode is a second opinion that helps you calm down and see one more reading than the one your anxiety handed you. It is not a verdict, and it is not a substitute for the sentence that resolves most of these situations in about nine seconds: “Hey — I want to make sure I’m reading you right. Are you upset with me?” If you find yourself running four different tools on one message, the message was never the problem. Ask her.
The question to ask before you paste her words into anything — including us
You are about to hand a private message from someone who trusts you to a company she has never heard of. That is worth ten seconds of thought, and almost nobody gives it any.
With a consumer chatbot: her words go to that company’s servers. Depending on your account settings, your conversations may be used to improve their models — there is usually a toggle for that, and most people have never opened it. Her name, if you paste it, goes too. So does everything else in the screenshot you cropped carelessly.
With us: we store your decodes. That is literally what makes your history work, and we’d rather say it than claim we store nothing while doing exactly what everyone else does. The message text is sent to Anthropic’s Claude to be analysed. We don’t sell your data. It’s all written down plainly in our privacy policy, including the parts that don’t flatter us.
Practical advice, whichever tool you use: strip her name. Paste the one message you’re stuck on, not the whole thread. And apply the test that actually matters — if she found out you’d done this, would you be able to explain yourself without flinching? Wanting to understand her is defensible. Most of what people do with these tools is exactly that. Just be honest with yourself about which one you’re doing.
What Decoded actually adds
Not intelligence. It runs on Claude, same family of model you’d be prompting yourself. We didn’t train anything, and we’re not going to pretend we did.
What it adds is consistency and one button. The same five-part output every single time. A tone flag on every decode — green, yellow or red — instead of a paragraph you have to interpret. A “don’t say this” you never have to remember to ask for. And no 200-word prompt to find while your hands are shaking. Here’s the same message, in Decoded’s format:
Most likely she does want you there, and she’s giving you a way out that she hopes you won’t take. It reads less like anger and more like someone bracing for you to choose the easy option. The literal reading — she genuinely doesn’t mind — is possible, but it’s the second-most-likely one given that she raised it at all.
Probably feeling low-priority rather than furious. The sentence protects her from having to ask you to want to be there.
Don’t: reply “ok, I won’t come then.” It answers her words and ignores her. If you genuinely can’t face it, say that you can’t and that you wish you could — the reason matters more than the outcome here.
An illustrative decode, written in the app’s voice. Decoded reads the message she actually sent you — what she means, the feeling underneath, a tone flag, what to send back, and the one thing not to say. Get early access →
Free is 5 decodes a day, no card, with one suggested reply and your last 3 decodes in history. Premium is $9.99 a month for unlimited decodes and three replies instead of one. Same model, same prompt, same analysis in both — the paid tier buys you volume, not a smarter answer. It isn’t out yet; iOS and Android, and we’ll email you the day it lands. Get early access if you want it.
And if you’d rather just keep the prompt and never hear from us again — genuinely, fine. It’s free for a reason. If you want to go deeper on how a general chatbot handles this, we wrote a straight answer to whether ChatGPT can analyse your text messages, and the guide to what “I’m fine” actually means is the one most people need at this hour.
Questions people also ask
Yes. ChatGPT can read an image, and it will pull the text out of a screenshot well enough to work with. Two cautions: it sometimes attributes messages to the wrong person when the bubbles are ambiguous, so check that it understood who said what before you trust the reading. And a screenshot of a whole thread contains far more of her private life than the one line you were worried about.
It depends entirely on what you do with the answer. Using it to understand her better before you reply is not so different from asking a friend. Using it to find the words that get you the outcome you want regardless of what she meant is a different thing, and you will know which one you are doing.
A useful test: would you be comfortable telling her you did this? If the answer is no, ask yourself why.
For most people the honest answer is: you may not need to. The prompt on this page is free, it works, and Decoded’s free tier gives you 5 decodes a day with no card.
What Decoded adds is consistency and speed — the same five-part output every time, a tone flag pinned to the same three levels — green, yellow, red — instead of a fresh adjective each time, a “don’t say this” you never have to ask for, and no 200-word prompt to remember at 1am. Premium ($9.99/month) removes the daily cap and gives you three suggested replies instead of one. It’s the same model and the same analysis, not a smarter answer.