Nearly everything I publish here is about ad accounts. This is about the other thing I build, and why it comes from the same instinct.
The message that gets typed, read back and deleted
Someone writes the thing they actually mean. They look at it. They decide the wording carries more risk than the point, and they send nothing, or they send a flatter version and regret that instead. I kept seeing that pattern in my own threads and in other people’s, and the more I looked at it the less it seemed like a writing problem.
In most hard messages the person already knows what they think. What they cannot judge is how it will land, and the cost of getting that wrong attaches to them personally. So the message stays in the notes app. The research on procrastination describes this as short term mood regulation, not laziness. Sirois and Pychyl found that putting something off is usually about not feeling the feeling now. And we are worse at reading our own tone than we think. Kruger, Epley, Parker and Ng showed people are confident their sarcasm and their warmth come through in text, and the reader gets it wrong far more often than the writer expects, which is the gap Subtext exists for.
What it does
Subtext is an app that reads the emotional tone of a message before you send it, then helps you say the same thing in your own voice. A message goes through six steps.
- You give it the message. A draft, a screenshot of the whole thread, or a voice note.
- It reads the whole conversation. Both sides of it, not your line alone, so it can tell you what the other person is most likely asking for, separate from everything wrapped around it.
- It names what your draft carries underneath. The hurt behind a short reply, the defense inside an apology, the guilt in a cancellation.
- Only then does it rewrite. It works inside your own phrasing instead of replacing it with a generic one.
- Each version gets a safe to send score. An estimate from 5 to 100 of how likely the message is to land the way you mean it.
- You decide. It is a read, not a guarantee. You know the relationship and the app does not, and it says so.
It runs on the phone where you were already writing, and on the web at web.subtext.it when the message is long enough to want a keyboard.
Why it reads before it rewrites
Every phone now rewrites messages.
| Tool | What it does with your draft | Reads the whole thread and says how your reply is likely to land |
|---|---|---|
| Apple Writing Tools and Gboard | Make it friendlier or more professional, and they are good at it | No |
| Grammarly’s tone detector | Tells you how a sentence sounds | No |
| Subtext | Reads the tone first, then rewrites (steps 3 and 4 above) | Yes |
The Apple and Grammarly entries follow what each company documents, not what I suspect.
I think that gap is deliberate. Telling someone their message sounds passive aggressive is a rude thing for a phone to do.
But naming the feeling is the part that helps. Lieberman and colleagues found that putting a feeling into words dampens the brain’s response to it, which is why the app analyzes first and offers wording second. A general chatbot tends to flatter you instead, because language models lean toward agreeing. Sharma and colleagues measured it, and OpenAI wrote up its own case in 2025. Cheng and colleagues then showed what it does to people. In three preregistered experiments, advice from a sycophantic model left participants less willing to repair a real conflict and more convinced they were right. A tool that agrees with you cannot tell you that your message is going to land badly. We built Subtext to disagree with you, politely, before the other person does.
What it will not do
If you want an app that flirts on your behalf, that category exists and we are not in it. If all you need today is a comma fixed, your keyboard has that covered. Subtext earns its place on the messages where the wording carries more risk than the argument, and it says plainly when a draft is fine as it is.
It also names what it runs on. Anthropic’s Claude does the reading, drafting and rewriting, including screenshots. Deepgram turns voice into text. OpenAI provides the optional voice that reads replies aloud. Google Firebase holds accounts and storage. We name all four in the privacy policy, because a product that asks you to paste in your most careful conversations does not get to be vague about where they go.
Why a growth person built it
The same instinct runs both of my businesses. In the consulting work I read an ad account before I spend a cent in it, because the account already knows things the brief does not. Subtext reads a message before you send it, for the same reason. In both cases the expensive mistake is acting on a guess when a read was available.
Two people in Berlin make it. Anton Licht writes the code. I do everything that involves words, including the app copy, the research articles and the fifty questions I answered about it. There is no team behind us and no agency. It ships in the same seventeen languages as this site. The app is free to download, but the AI features need a paid subscription through the App Store or Google Play, with the price shown in your own currency before you commit.
Where to look
The Subtext AI writing assistant is at subtext.it. The research it stands on, more than forty articles with their sources, is on the Subtext blog. The reviews it has earned, named and verbatim, are at subtext.it/reviews. And the long version of everything above, fifty questions I answered about it, is here on this site.